Citation
E-WOM communication in online consumer review sites and social networking services

Material Information

Title:
E-WOM communication in online consumer review sites and social networking services
Creator:
Aghakhani, Navid ( author )
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file (101 pages) : ;

Thesis/Dissertation Information

Degree:
Doctorate ( Doctor of philosophy)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
Department of Computer Science and Engineering, CU Denver
Degree Disciplines:
Computer science and information systems
Committee Chair:
Karimi, Jahangir
Committee Members:
Gregg, Dawn
Oh, Onook
Ra, Ilkyeun

Subjects

Subjects / Keywords:
Data mining ( lcsh )
Social psychology ( lcsh )
Consumers -- Behavior ( lcsh )
Marketing -- Psychological aspects ( lcsh )
Marketing -- Social aspects ( lcsh )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Review:
Electronic Word of Mouth (eWOM) communication has become an important research topic both for marketing and information systems scholars. To explore this topic, this dissertation is comprised of two essays focusing on adoption of eWOM in Social Networking Services (SNS) and credibility of eWOM in online consumer review sites. In the first essay, I investigate the antecedents of adoption of eWOM on Facebook and I suggest two types of eWOM on Facebook. The first type of eWOM is explicit eWOM, which is delivered through written text format. The second type of eWOM is implicit eWOM, which is specific to SNS, and delivered through non-written text cues such as “Like” and “Check-in”. To investigate the adoption of explicit and implicit eWOM, I adopted the theoretical lens of Elaboration Likelihood Model (ELM) and Affect-As-Information Theory. Data for this study is collected through a survey of 202 students. In the second essay, I examine drivers of review usefulness in online consumer review sites. Review usefulness, as the flip side of review credibility, has been the most important driver of eWOM adoption and consumer purchase decision making in online consumer review sites. For this exploration, I adopted the theoretical lens of ELM and collected 2,634 online consumer reviews and the metadata associated with them from Yelp. These two essays contribute to existing literature in eWOM communication in two important ways. The first essay elucidates how the interplay between central and peripheral cues of ELM and Affect-As-Information theory explain the adoption of explicit and implicit eWOM on Facebook. The second essay contributes to ELM by crafting and operationalizing a new variable of review consistency, which measures the consistency between central and peripheral cues of ELM in the online review context. The results of these two essays have important managerial implications, particularly, for social media marketing strategy and improving recommender systems in online consumer review sites.
Thesis:
Thesis (Ph.D..)--University of Colorado Denver.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: Adobe Reader.
General Note:
n3p
Restriction:
Embargo ended 05/09/2019
Statement of Responsibility:
by Navid Aghakhani

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
995175747 ( OCLC )
ocn995175747

Downloads

This item has the following downloads:


Full Text
E-WOM COMMUNICATION IN ONLINE CONSUMER REVIEW SITES AND SOCIAL
NETWORKING SERVICES by
NAVID AGHAKHANI B.S., Shahid Beheshti Universit, 2007 M.S., University of Malaya, 2011
A thesis submitted to Faculty of the Graduate School of the The University of Colorado in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Computer Science and Information Systems Program
2017


This thesis for the Doctor of Philosophy degree by
Navid Aghakhani
has been approved for the
Computer Science and Information Systems Program
by
Jahangir Karimi, Chair Dawn Gregg, Advisor Onook Oh Ilkyeun Ra
Date: May 13, 2017
11


Aghakhani, Navid (Ph.D., Computer Science and Information Systems)
E-WOM Communication in Online Consumer Review Sites and Social Networking Services
Thesis directed by Professor Dawn Gregg
ABSTRACT
Electronic Word of Mouth (eWOM) communication has become an important research topic both for marketing and information systems scholars. To explore this topic, this dissertation is comprised of two essays focusing on adoption of eWOM in Social Networking Services (SNS) and credibility of eWOM in online consumer review sites. In the first essay, I investigate the antecedents of adoption of eWOM on Facebook and I suggest two types of eWOM on Facebook. The first type of eWOM is explicit eWOM, which is delivered through written text format. The second type of eWOM is implicit eWOM, which is specific to SNS, and delivered through non-written text cues such as “Like” and “Check-in”. To investigate the adoption of explicit and implicit eWOM, I adopted the theoretical lens of Elaboration Likelihood Model (ELM) and Affect-As-Information Theory. Data for this study is collected through a survey of 202 students. In the second essay, I examine drivers of review usefulness in online consumer review sites. Review usefulness, as the flip side of review credibility, has been the most important driver of eWOM adoption and consumer purchase decision making in online consumer review sites. For this exploration, I adopted the theoretical lens of ELM and collected 2,634 online consumer reviews and the metadata associated with them from Yelp. These two essays contribute to existing literature in eWOM communication in two important ways. The first essay elucidates how the interplay between central and peripheral cues of ELM and Affect-As-Information theory explain the adoption of explicit and implicit eWOM on Facebook. The second essay contributes to ELM by crafting and operationalizing a new variable of review consistency, which measures the consistency between central and peripheral cues of ELM in the online review context. The results


of these two essays have important managerial implications, particularly, for social media marketing strategy and improving recommender systems in online consumer review sites.
The form and content of this abstract are approved. I recommend its publication.
Approved: Dawn Gregg
IV


This dissertation is dedicated to my parents for their love, endless support and encouragement. Also, this dissertation is dedicated to my brother and sister, Nima and Anahita, who have been a
great source of motivation and inspiration.
Finally, this dissertation is dedicated to all those who believe in the richness of learning.
v


ACKNOWLEDGMENTS
First and foremost, I wish to thank my committee members who were more than generous with their expertise and precious time. Special thanks to Dr. Dawn Gregg, Dr. Jahangir Karimi and Dr. Onook Oh for their countless hours of reflecting, reading, encouraging, and most of all patience throughout the entire process. Also, I would like to thank Dr. Ilkyeun Ra from the computer science department for serving on my committee. I wish to thank my friends, fellow doctoral students, specially Gisella Bassani, for their help and support in this journey. Finally, I would like to thank my parents for allowing me to realize my own potential. All the support they have provided me over the years was the greatest gift anyone has ever given me.
vi


TABLE OF CONTENTS
I. DISSERTATION OVERVIEW............................................................ 1
Adoption of eWOM in SNS...........................................................2
Credibility of eWOM in online consumer review sites...............................4
II. A UNIFIED MODEL FOR ADOPTION OF EWOM ON FACEBOOK..................................6
Abstract..........................................................................6
Introduction......................................................................6
Theoretical Framework.............................................................9
Elaboration Likelihood Model.................................................11
Affect-As-Information Theory.................................................14
Hypotheses Development...........................................................16
Research Methodology.............................................................23
Measurement of Constructs....................................................23
Data Collection..............................................................24
Sample Profile...............................................................25
Data Analysis................................................................25
Discussion.......................................................................28
Implications.....................................................................30
Theoretical Implication......................................................30
Practical Implication........................................................33
Limitations and Future Research Direction........................................35
Conclusion.......................................................................36
vii


III. THE EFFECT OF REVIEW CONSISTENCY ON USEFULNESS OF ONLINE
REVIEWS: EVIDENCE FROM REVIEWS IN SERVICE CONTEXT................................43
Abstract.........................................................................43
Introduction.....................................................................44
Theoretical Foundation...........................................................49
Elaboration Likelihood Model.................................................49
ELM and Review Usefulness....................................................50
ELM’s Central Cue............................................................51
ELM’s Peripheral Cues........................................................52
Review Consistency...........................................................53
Research Methodology.............................................................55
Data Collection..............................................................55
Measures.....................................................................55
Empirical Models.............................................................60
Discussion.......................................................................63
Post-hoc Analysis............................................................65
Theoretical and Practical Contributions..........................................68
Theoretical Contributions....................................................68
Practical Contributions......................................................69
Limitations and Future Research..................................................69
Concluding Remarks...............................................................70
REFERENCES...........................................................................78
APPENDIX.............................................................................89
viii


Appendix A: Summary of Review Helpfulness Studies............................89
Appendix B: Negative Binomial Regression.....................................90
IX


LIST OF TABLES
Table II-1. Measurement Items.........................................................38
Table II-2. Reliability and Convergent Validity.......................................39
Table II-3. Discriminant Validity.....................................................40
Table II-4. Cross Loadings............................................................40
Table II-5. Common Method Bias........................................................41
Table III-1. Descriptive Statistics...................................................74
Table TTT-2 Correlation Matrix........................................................75
Table III-3. Categories of Useful Votes...............................................75
Table III-4. Econometric Analyses, Ordered Logit......................................76
Table III-5. Comparing Successive Models..............................................77
Table III-6. Post-hoc Analysis........................................................77
x


LIST OF FIGURES
Figure II-1. Research Model............................................................38
Figure II-2. Empirical Results.........................................................42
Figure II-3. Social Advertising in Amazon..............................................42
Figure III-1. An example of a Review on Yelp...........................................72
Figure TTT-2 Research Model............................................................72
Figure III-3. Supervised Measure of the Review Consistency.............................73
Figure III-4. The Distribution of Rating Based on Number of Reviews....................73
xi


CHAPTER I.
DISSERTATION OVERVIEW
Research on eWOM has attracted scholars from various disciplines, especially marketing and Information Systems. This is mainly due to the significant impact of users’ recommendations on potential buyers’ attitude toward brands or purchase intention. The term eWOM has different definitions, however, in this dissertation I rely on the definition by Henning-Thureau et al. (2004), “any positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of people and institutions via the Internet” (Hennig-Thurau et al.,2004, p. 39).
Studies on impact of eWOM communication is classified into two levels: market-level analysis and individual-level analysis (Cheung and Thadani, 2012; Lee and Lee, 2009). The focus of market-level analysis is on the effect of eWOM on product sales (Chevalier and Mayzlin 2006; Raguseo and Vitari 2017). At the individual level, however, the focus of research is on personal influence of the eWOM, where, the sender of the message can affect consumer purchase decision (Cheung et al. 2012; Cheung et al. 2009). This dissertation falls in the individual level effect of eWOM communication research.
The effect of eWOM communication both at market-level and individual level has been studied in four main platforms: online discussion forums (Schindler and Bickart 2012; Zhang and Watts 2008), online consumer review sites (Awad and Ragowsky 2008; Baek et al. 2012; Kuan et al. 2015), blogs (Chu and Kamal 2008) and online shopping sites (Gupta and Harris 2010; Park et al. 2007). Online consumer review websites are the most common platforms to study eWOM. Major published studies in this stream tend to investigate the drivers of eWOM credibility (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009; Kuan
1


et al. 2015). This is largely due to users’ intention to read recommendations about products from actual buyers’ perspective or to express their own opinions about products that they have already purchased.
Adoption of eWOM in SNS
Social Network Services (SNS) have recently been identified as perfect platforms for spreading eWOM and brand related information (Chu and Kim 2011; Fang 2014). Collaborative and social characteristics of SNS has enabled their users to show high level of social presence by connecting to other users, exchanging information and opinions (Kaplan and Haenlein 2010). The desire to establish and maintain social relationships also enables eWOM behavior among SNS users (Chu and Kim 2011). By sharing useful products and brand related information users can help their friends (connections) with their purchase-related decisions (Fang 2014). The increasing use of SNS can also facilitate close relationships among organizations and consumers as a component of an Integrated Marketing Communication (IMC) (Mangold and Faulds 2009). Users can search for unbiased product information, share their own consumption-related advice and display interpersonal comments publicly (Fang 2014; Luam et al. 2015). The opinion passing behavior that exists in SNS gives an important, yet overlooked, dimension to SNS in comparison to other platforms of eWOM, which merely focus on opinion seeking and opinion giving as the only two ways of eWOM communication (Norman and Russell 2006). The online passing of information enabled by SNS facilitates the flow of information regardless of geographical and temporal constraints (Chu and Kim 2011). Consumers can quickly exchange brand related information with a few clicks with their connections as well as global audience who share common interests (Norman and Russell 2006). As a result, SNS as a platform for online branding and
2


advertising has undergone tremendous growth and advertising spending on SNS in the US is expected to reach $2.6 billion by 2012 (Chu and Kim 2011).
Among existing SNS, Facebook is the most popular, attracting various corporations to open their own fan pages to advertise their brands. As such, Facebook is widely considered a platform of choice for generating and spreading recommendations about specific brands or products and for engaging in eWOM activities (Chu and Kim, 2011). In Facebook, these recommendations have been categorized into explicit and implicit types (Ebermann et al. 2011). The option to use direct and indirect communication channels in Facebook is similar to the categorization of recommender systems based on explicit and implicit types. In user profiles, information can be provided via status messages or in predefined categories such as likes and interests. The main goal of information provided in a user’s profile is to present the user and his/her preferences (Liu 2007). Although the major goal of a user’s profile is not to recommend something, it can have recommendation effects on the other users who read it, because it may refer to the products or services the user likes (Luarn et al. 2015). In other words, while profile information is not directed specifically at other users, it might have a potential unintended recommendation effect, which is considered implicit recommendation. Conversely, explicit recommendations are intentionally provided from one SNS user to other users. Such recommendations may be given through direct communication channels such as webmail-like messaging within SNS or as a direct response to recommendation requests in status messages (Ebermann et al. 2011). Given the emergence of SNS as a new platform of eWOM communication and by considering the difference between explicit and implicit eWOM, the first essay of this dissertation will explore the
3


drivers of eWOM adoption through the theoretical lens of Elaboration Likelihood Model and Affect-as-information theory.
Credibility of eWOM in online consumer review sites
Trust in the online environment has been an important research topic. Especially, in the space of online transactions where uncertainty is prevalent, users’ trust in online information has been shown to be an important driver of the effective marketplace. For instance, regarding eBay, prior literature has focused on the trust issues between buyers and sellers. These trust issues stem from the inherent information asymmetry problem in the online marketplace, where buyers have more information about sellers, thus increasing the uncertainty in online transactions. Prior studies show that the implementation of an online feedback system, where consumers can share their opinion about sellers enhance trust in eBay, thereby, increasing the price premium (Gefen et al. 2008; Pavlou 2003; Pavlou and Dimoka 2006; Pavlou and Gefen 2004).
Another aspect of trust in the online environment is related to credibility of users’ generated content in online consumer sites. The overwhelming volume of online reviews presents challenges for consumers seeking relevant and trustworthy reviews (Baek et al. 2012; Kuan et al. 2015; Mudambi and Schuff 2010). To address this problem, online consumer review websites such as Yelp and Amazon have implemented online social voting systems. An online social voting system is a technological component that allows consumers to diagnose the credibility of online reviews by voting for its usefulness or helpfulness (Baek et al. 2012). Understanding the factors that impact the helpfulness/usefulness of online reviews is an important question for both academics and practitioners because online review quality is directly related to consumers’ level of trust in the online marketplace. In addition,
4


from a theoretical perspective, review usefulness (helpfulness), as the proxy for review credibility, is known to be the most important driver for consumers to adopt eWOM (Cheung et al. 2012; Cheung et al. 2009). Other studies have reported that online reviews that are perceived as useful exert a stronger influence on an online user’s decision to conduct online transactions than less useful reviews (Baek et al. 2012; Yin et al. 2014a). From a practical perspective, online retailers that display helpful reviews gain a strategic advantage in consumer attention and “stickiness” (Connors et al. 2011). Reflecting the important role of social voting systems in the online marketplace, Amazon added an extra $2.7 billion to its revenue by implementing its social voting system and asking consumers to rate the helpfulness of its online reviews (Spool 2009).
To understand the phenomena of credibility of online reviews, the second essay of this dissertation focus on the drivers of review usefulness on Yelp. This essay draws on the burgeoning body of research that emphasizes how the content of online reviews affects reviews’ usefulness.
5


CHAPTER II.
A UNIFIED MODEL FOR ADOPTION OF EWOM ON FACEBOOK
Abstract
Electronic word of mouth (eWOM) is a concept that has gained increased attention from both practitioners and academia. Its importance lies in its simplicity and yet its profound impact on customers’ attitudes toward specific brands or goods, thus affecting customers’ loyalty and purchase behaviors. Although social network services (SNS) have emerged as a new platform for eWOM communication, less attention has been paid in the literature to the adoption of eWOM in SNS. Using the elaboration likelihood model (ELM) and the affect-as-information theory, this study identifies factors that affect the adoption of eWOM on Facebook. We identify product-related attributes of a review, source credibility, peer image, and tie strength as theoretically important variables in our study, and we examine their effect on cognitive and affective attitude.
We find that the eWOM type (explicit vs. implicit) moderates the effects of cognitive and affective attitude on the adoption of eWOM. We further find that the effect of cognitive attitude on eWOM adoption is higher when the eWOM is explicit, while the effect of affective attitude is higher when eWOM is implicit.
Keywords: Electronic Word of Mouth (eWOM), eWOM Adoption, eWOM type, cognitive attitude, affective attitude
Introduction
Advances in information technology in general and the emergence of web 2.0, in particular, have opened new avenues of research through which scholars can study the effect of eWOM on consumers’ product and service judgments (Lee and Youn 2009). eWOM has
6


been defined as “any positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of people and institutions via the Internet” (Hennig-Thurau et al. 2004). Most of the previous studies on adoption of eWOM were done on conventional platforms for eWOM communication, such as online consumer review websites (Awad and Ragowsky 2008; Cheung et al. 2009; Lee and Lee 2009; Lee and Youn 2009; Park et al. 2007), blogs (Chu and Kamal 2008; Riegner 2007), and online shopping sites (Gupta and Harris 2010; Lee et al. 2008; Pan and Chiou 2011). However, the emergence of social networking services (SNS) has brought significant attention from the commercial marketplace, with global advertisement spending on SNS predicted to exceed $35 billion in 2015, and has opened a new channel for generating and disseminating eWOM (LePage 2015). Among existing SNS, Facebook is currently the most popular (eMarketer.com 2014), accounting for 50% of all social referrals and 64% of social revenue (Cooper 2015). As such, Facebook is generally considered to be the platform of choice for generating, spreading, and encountering eWOM (Chu and Kim 2011).
While considerable attention has been paid to the communication and adoption of eWOM on conventional eWOM platforms, less attention has been paid to the adoption of eWOM on Facebook. This study is particularly motivated by the failure of prior studies to consider the influence of different eWOM types on the drivers of eWOM adoption on Facebook. Past eWOM adoption studies on both Facebook (Chu and Kim 2011; Fang 2014) and conventional eWOM platforms (Cheung et al. 2012; Cheung et al. 2009) considered only eWOM generated through written text. However, non-textual information in a Facebook user’s profile, such as “Likes” and “Check-ins,” also has the potential to influence the decision making of the user’s peers (Keenan and Shiri 2009; Luam et al. 2015; Okazaki
7


2009). In this study, we use the term “implicit eWOM” to refer to eWOM that is delivered through non-textual cues such as the “Like” and “Check-in” buttons, and we use the term “explicit eWOM” to refer to eWOM proliferated through written text. By considering the notion that Facebook provides a forum for communication of both explicit and implicit eWOM, and by using other theoretically important variables, three research questions are addressed: (1) How do the cognitive and affective attitudes play a role in the eWOM adoption process? (2) what are important factors that impact the cognitive and affective attitudes for explicit and implicit eWOM adoption? (3) do eWOM type moderates the impact of cognitive and affective attitudes on eWOM adoption. To address the above questions, we consider eWOM through the theoretical lens of the elaboration likelihood model (ELM) (Petty and Cacioppo 1986) and the affect-as-information theory (Forgas 1995; Forgas and George 2001). We then identify the factors that have a direct impact on cognitive and affective attitudes by investigating the role of product-related attributes of a review, source credibility, peer image building, and tie strength in our research model. Next, we investigate whether eWOM type moderates the effects of cognitive and affective attitudes on eWOM adoption. This study, therefore, contributes to the existing literature eWOM adoption in several ways. First, it distinguishes between explicit and implicit eWOM by suggesting that implicit eWOM is more unique to Facebook. Second, it integrates ELM and the affect-as-information theory to explore how the combination of affective and cognitive attitudes explain eWOM adoption process. Third, by empirically examining the moderating impact of eWOM type on the effects of cognitive and affective attitudes on the eWOM adoption, this work contributes to the extant literature in eWOM adoption process.
8


Below we provide the theoretical framework for this study followed by hypotheses development. In the methodology section, we describe an empirical study designed to test the proposed model using a sample data from students at two large public universities. We then discuss the measurement scales for the constructs in our research model. Next, we present the results of our empirical findings, the study’s theoretical and practical implications, limitations and conclusions.
Theoretical Framework
Facebook differs from conventional eWOM communication platforms in three important ways. First, consumers use online retailer websites, forums, and blogs specifically for the purpose of seeking and giving opinions about products and services (Lee and Lee 2009; Sun et al. 2006). In other words, the focus of these platforms is on the eWOM itself. In contrast, the primary drivers of SNS use are social presence and interpersonal connections among users (Fang 2014). Second, the eWOM content on Facebook is pushed toward users by appearing on their wall or newsfeed, and the source of this eWOM is not anonymous but rather one of the user’s Facebook friends (Chu and Kim 2011; Fang 2014; Luarn et al. 2015). Third, prior eWOM communication studies have considered online reviews (eWOM generated in a written text format) as the main form of eWOM (Cheung et al. 2012; Cheung et al. 2008; Cheung et al. 2009; Chu and Kim 2011; Fang 2014). On Facebook, users can explicitly state their opinions about products or services in a written text format, for example, by publishing a review in the form of a Facebook status update, by posting comments, or by using Facebook’s direct message mechanism. However, Facebook users can also show their interest in products and services by using non-textual means, such as by using the “Like” or “Check-in” buttons to share specific information about a brand on their profiles (such as
9


promotions, new products, etc.). Although the main goal of the information provided in a user’s profile is to present the users’ preferences (Liu 2007), it can also have the effect of a recommendation for other members (Ebermann et al. 2011). For instance, location-based services such as Facebook’s “Check-in” button allow users to share information on their timeline about places they have visited. A user’s timeline thus shows his/her consumption behavior and interest in certain products and services, and this information has the potential to influence the decision making of other SNS users (Keenan and Shiri 2009; Luarn et al. 2015; Okazaki 2009). It is now a common practice for businesses to seek to persuade consumers to “Like” or share their brands on Facebook, as this action can affect other users’ decision-making process. A 2012 report shows that almost 90% of Facebook users have “Liked” at least one brand on Facebook, and 69% have “Liked” a brand because a friend in their network had also liked it (PurelyBranded.com 2012).
Given the differences between explicit and implicit eWOM, and in view of the social-emotional relationships among Facebook users, in this study we use the lens of informational and relational influences to consider the process of eWOM adoption on Facebook (Cheung et al. 2009; Hennig-Thurau et al. 2004; Shih et al. 2013). Prior eWOM communication studies have considered the informational influence of eWOM as the effect of the online review content, such as the elaboration of a product’s performance or the strength of the review’s argument, on the consumers’ cognitive attitude (Baek et al. 2012; Cheung et al. 2012;
Cheung et al. 2008; Cheung et al. 2009; Kuan et al. 2015). Consistent with these prior studies, one of the paths for eWOM adoption on Facebook is the informational influence of the content of explicit eWOM on Facebook users’ cognitive attitude. In addition to the informational influences of eWOM, recent studies have also examined the relational
10


influences and social bond characteristics of a platform in predicting eWOM behavior (Fang 2014; Lee et al. 2012; Shih et al. 2013). This is particularly relevant in the case of Facebook, where sharing of moods, feelings, thoughts, and locations is considered as a form of social interaction; thus, factors related to relational influences (i.e. socio-emotional factors) play an important role in the adoption of eWOM on Facebook (Keenan and Shiri 2009; Luarn et al. 2015; Okazaki 2009). Moreover, product and brand information shared through consumers’ “Likes” and “Check-ins” (i.e., implicit eWOM) can be considered as a type of advertisement for those products and brands, and numerous studies have identified the affective attitude of consumers toward advertisements as the main driver for adoption of such ads (Edell and Burke 1984; Huang et al. 2013; Leung et al. 2015; MacKenzie et al. 1986; Shimp 1981). Building on these prior studies, in this study we consider cognitive attitude and affective attitude as the two paths for eWOM adoption on Facebook, and we consider the informational and relational (i.e., the socio-emotional) influence factors that affect the cognitive and affective paths. Furthermore, we examine the suitability of the ELM and the affect-as-information theory for exploring the drivers of cognitive and affective attitude, respectively.
Elaboration Likelihood Model
The Elaboration Likelihood Model explains that cognitive attitude change among people can be triggered by two routes of influence, the central route and the peripheral route. The primary difference between these routes lies in the amount of thoughtful information processing that is required of the individual subject (Bhattacheijee and Sanford 2006). The central route requires an individual to think critically about the issues involved in the argument of a message and to inspect the relative facts and relevance of the message before
11


making a conversant judgment about the message. Therefore, processing a message through the central route demands a high cognitive effort. Perception changes that arise from the central route are more stable and long-lasting, as they are based on careful consideration of message cues and are therefore more predictive of long-term behavior. Conversely, the peripheral route involves the processing of heuristic cues (Angst and Agarwal 2009; Bhattacheijee and Sanford 2006; Petty and Cacioppo 1986); it merely requires an individual’s association with positive and negative cues in the message argument. As such, processing a message through the peripheral route entails less cognitive effort, and alterations caused by peripheral route influences are less persistent and predictive of long-term behavior. Despite the distinctions between these two routes, in practice, people typically evaluate a message through a modest level of employing both routes (Sussman and Siegal 2003).
The suitability of the ELM for understanding how people process messages that are intended to be persuasive has been examined previously in information systems (IS) research. Bhattacherjee and Sanford (2006) used the ELM to examine the influence processes for information technology acceptance. Sussman and Siegal (2003) employed the ELM in a non-experimental setting to study knowledge adoption via electronic mail by consultants at a public accounting firm. Angst and Agarwal (2009) used the theoretical lens of the ELM to study the adoption of electronic health records in an experimental setting. The ELM has also been used in past studies as the theoretical lens for eWOM communication and elaboration. For example, the ELM has been used in studies on the effect of negative online consumer reviews on consumers’ product attitudes (Lee et al. 2008), the effect of eWOM on consumers’ choice of product (Gupta and Harris 2010), perceived blogger credibility and the
12


impact of eWOM argument quality on brand attitudes (Gupta and Harris 2010), and the credibility of eWOM in online consumer review websites (Cheung et al. 2012).
Based on these prior studies, we believe the ELM provides an appropriate theoretical lens to understand the drivers of cognitive attitude change in the context of eWOM adoption on Facebook. Cognitive attitude has been defined as the degree to which an individual develops beliefs relating to the attitude object (Shih et al. 2013). The notion of the attitude object varies based on the context of the study. For instance, past studies have examined perceived usefulness in the context of the adoption of information technology (Adams et al. 1992; Davis 1989) and trust in the context of the online marketplace (Gefen et al. 2008; Pavlou and Gefen 2004). In the context of online consumer reviews, perceived credibility of online reviews has been examined as the main driver of the adoption of eWOM (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009). In the context of adoption of eWOM on Facebook, however, we believe it is necessary to consider the differences between Facebook and online consumer review sites in conceptualizing the attitude construct. This perspective is also important in sense that the major role of Facebook is hedonic, meaning that people tend to use Facebook for the purpose of social presence in the form of online interpersonal interactions and relationships (Kaplan and Haenlein 2010). This perspective is also consistent with the study by (Shih et al. 2013), wherein they examined the drivers of eWOM intention in an online forum. They argued that, because the social interaction among members was one of the most important and integral aspects of the forum, the cognitive attitude construct reflected whether the members believed the use of the online forum to access online reviews was wise, beneficial, and valuable. Consistently, we adopt the same conceptualization of the cognitive attitude in this study, and we define cognitive attitude as the degree to which
13


Facebook users believe accessing online reviews through Facebook is wise, beneficial, and valuable.
Building on the ELM, we consider the central and peripheral routes as the antecedents of cognitive attitude. In this study, we consider the product-related attributes of the eWOM as the factors that relate to the central route of the ELM. Product-related attributes of the eWOM can be defined as the degree to which an eWOM posted on Facebook reflects on the attributes of a product. Since source credibility is considered the most important heuristic driver of eWOM adoption in online consumer review platforms (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009), therefore, it is considered as one of factors related to the peripheral route of eWOM adoption. Source credibility is defined as the degree to which a recipient of eWOM on Facebook believes the sender of the eWOM (his/her friend) has knowledge, trustworthiness, credibility, and expertise regarding the product/service. In addition, tie strength and peer image are considered as social relationship variables and part of the peripheral route of eWOM adoption in our model. They have also been studied in prior studies of eWOM communication via SNS (Casteleyn et al. 2009; Chu and Kim 2011; Hansen and Lee 2013; Lin and Utz 2015; Luarn et al. 2015; Svensson 2011). Tie strength is defined as degree to which a Facebook user believes that his/her friends are close to him/her. Peer image is defined as the degree to which a Facebook user believes that his/her peer uses Facebook to shape an impression of himself/herself.
Affect-As-Information Theory
According to the affect-as-information theory, a person’s affective attitude can impact his/her assessment of the consequences of potential actions and decisions (Zadra and Clore 2011). Depending on the individual’s personality and the type of judgment being used, the
14


person’s emotional attitude can lead to doubt or confidence in his/her evaluation of a target object. People also utilize their emotional attitude as a shortcut for evaluating a target in terms of social behavior. This process is very common when quick judgment or heuristic processing is required (Forgas 1995). Prior studies in the areas of management and marketing have investigated the role of affective attitude in consumers’ decision-making processes. For instance, the affect-as-information theory has been utilized to investigate the impact of employees’ affect and mood on their judgments and decision making in an organizational setting (Forgas and George 2001). It has also been used to examine the impact of a website’s atmosphere on users’ approach toward the website (Eroglu et al. 2003). Although affective issues are often overlooked within the IS field (Yin et al. 2014a), recent studies have begun to incorporate affect in their conceptual framework. This is especially true in the context of eWOM adoption, where the role of affective cues embodied in a review’s text has been examined as a driver of the perceived helpfulness of online reviews (Salehan et al. 2015; Yin et al. 2014a). The affect-as-information theory has also been utilized to understand the role of the emotional status of senders and receivers of eWOM in the eWOM adoption process (Soderlund and Rosengren 2007). In a similar vein, in the context of SNS, (Fang 2014) found that the affective attitude of Facebook users when reading eWOM can positively influence the eWOM adoption on Facebook.
Prior studies indicate that the arousal dimension of the affective attitude has a strong influence on information processing by individuals (Corson and Verrier 2007; Vogt et al. 2008). Arousal reflects the degree to which an individual is excited and stimulated. If someone is aroused, then he/she is likely to make a more positive judgment of the target task or object. Arousal can impact individuals’ perception of an issue and can drive them to give a
15


prompt response (Storbeck and Clore 2008). The study by Fang (2015) shows that the level of affect (arousal) resulting from reading online reviews is one of the drivers of eWOM adoption on Facebook. In addition to reading online reviews (explicit eWOM), Facebook users are often exposed to information about brands and products via their friends’ “Likes” and “Check-ins” (i.e., implicit eWOM), and those “Likes” and “Check-ins” may influence consumers’ decision-making process (Keenan and Shiri 2009; Luam et al. 2015; Okazaki 2009). Thus, in line with the affect-as-information theory, the arousal triggered by exploring “Likes” and “Check-ins” can serve as a source of information for the implicit eWOM adoption. Given the fact that Facebook contains both explicit and implicit eWOM, the affective attitude is conceptualized as the degree of arousal resulting from exposure to product information through friends’ explicit and implicit eWOM. Since prior studies have also highlighted the role of social relationship variables as the driver of affective attitude (Fang 2014; Shih et al. 2013), in our research model, the tie strength and peer image building are also included as the drivers of affective attitude.
Hypotheses Development
The effect of content of online reviews on the eWOM adoption has been examined in prior studies in online consumer review platforms (Ghose and Ipeirotis 2011; Qiu et al.
2012; Schindler and Bickart 2012). Schindler and Bickart (2012) found that the amount of product-descriptive statements in a review was associated with review usefulness. In a similar vein, Qiu et al. (2012) found that product-related attributes of a review enabled consumers to obtain information about characteristics of products and thus helped the consumers with their purchase decisions. Conversely, reviews containing non-product-related information revealed little information about products, and consumers found those reviews to
16


be less diagnostic and less credible. Since the review text plays an important role in the formation of individuals’ attitude toward eWOM in online consumer review platforms, and with regard to notion that the explicit eWOM adoption on Facebook is highly associated with the cognitive path, one would expect that product related attributes have a positive impact on the degree to which an individual develops attitudes about eWOM adoption on Facebook. Thus, we propose the following:
Hi: Product-related attributes of a review have a positive effect on the cognitive
attitude.
Online reviews with a high source credibility tend to support eWOM acceptance (Cheung et al. 2012; Cheung et al. 2009). Prior studies based on reviews collected from Amazon have found that source credibility has a positive impact on the helpfulness of online reviews. Forman et al. (2008) reported, for example, that revealing the identity of a reviewer had a positive impact on the helpfulness of reviews. More recent studies (Baek et al. 2012; Kuan et al. 2015) have demonstrated that platform-based signals of source credibility, such as the top reviewer badge on Amazon, have a positive impact on a review’s perceived helpfulness.
Despite the fact that source credibility in online consumer review websites is only considered as a “virtual credential” for the eWOM source (Cheung et al. 2009), it is still an important indicator of information credibility (Wathen and Burkell 2002). On Facebook, however, because of the rich social interactions and interpersonal relationships among members, the perception of a member’s credibility is more likely to be formed based on that member’s interactions with other members, rather than on virtual credentials (Chu and Kim 2011). Facebook friends encompass both close ties, such as immediate family, relatives, and friends that meet face-to-face on a regular basis, as well as more distant acquaintances that
17


may only interact through virtual channels. Becoming friends on Facebook allows users to have access to each other’s personal information and content notifications, such as status updates, comments, photographs, visited places, promotions, graduations, and so forth. Thus, Facebook enables users to maintain social relationships and to establish trust, which may extend to other contacts as well (Chu and Kim 2011). It has been found that the eWOM provided by friends is perceived as more credible than that from anonymous or personally unknown sources because these contacts are embedded in the consumers’ personal network (Chu and Kim 2011). On Facebook, therefore, people may consider recommendations from friends or classmates as more credible (Chu and Choi 2011). This makes SNS a vital source of product information for consumers that greatly enables the communication of eWOM. Therefore, it is no wonder that marketers have invested substantial resources in setting up brand profiles on SNS so as to engage consumers with their brand and to spread positive eWOM through SNS members (Jansen et al. 2009). Thus, because of the importance of source credibility in explicit eWOM adoption, we propose the following:
H2: Source credibility has a positive effect on the cognitive attitude.
Social identity and relationships among users have been identified as the main focus of SNS, particularly Facebook (Svensson 2011). As social image is an asset that Facebook users can use to maintain and enhance their status within their network (Luam et al. 2015), some users seek to increase their social identity and present an ideal picture of themselves, rather than the reality (Casteleyn et al. 2009). Image building is thus used by people who aim to publish content that matches the ideal image of themselves that they wish to create (Kaplan and Haenlein 2010). Presenting an idyllic image of oneself on Facebook can be done by various means. It can happen either explicitly through posting status updates or implicitly by the use of “Check-ins” or “Likes” (Keenan and Shiri 2009). For instance, past studies
18


show that Facebook users share their location information as an indirect way of enhancing their self-presentation and social image so that they appear more appealing to others in the social network. Thus, it is less likely that people will recommend a product that they believe will damage their social image (Luam et al. 2015).
We believe the effect of peer image building behavior on eWOM adoption on Facebook can be explained based on Facebook users’ underlying inferences about the motives of the friend who posted the eWOM. eWOM readers’ perception about the reviewers’ motives for recommending a product or service can be classified as either internal (i.e., self-serving reasons) or external (i.e., product-related reasons) (Lee and Youn 2009). If it is perceived that the reviewers’ motives are related to a product (external), consumers will perceive the review as helpful. On the other hand, if the inferred motives are self-serving (internal), consumers will discount the review (Sen and Lerman 2007). Accordingly, one may recommend a product or service for internal reasons (i.e., image building) or external reasons (i.e., product related). A qualitative study by (Svensson 2011) shows that eWOM can be ineffective if it is perceived to be communicated for internal reasons. If the eWOM is “too good” and thus is perceived only to communicate the desired personality of the sender, this reduces the sender’s credibility; consequently, it makes the message less reliable and may even cause it to be rejected. Therefore, image building behavior of a Facebook member reduces his/her credibility among his/her peers on Facebook and makes the member’s eWOM recommendation less likely to be accepted. The formation of consumers’ cognitive attitude about the suitability of Facebook for accessing eWOM may thus be formed in part based on their peers’ image building behavior. In addition, prior studies have also shown that the social attractiveness of individuals makes them more persuasive in general (McCroskey
19


et al. 1974). On Facebook, in particular, this stimulates the receivers’ arousal and makes them more excited to review the eWOM (Fang 2014). Consistent with these findings, we believe that the image building behavior of an individual on Facebook leads other members to be less excited about checking their “Likes” or “Check-ins.” Thus, peer image building on Facebook influences the eWOM adoption by having a negative effect on the receivers’ cognitive and affective attitudes. Therefore, we propose the following:
H3. Peer image building has a negative effect on the cognitive attitude.
H4: Peer image building has a negative effect on affective attitude.
In offline settings, the degree of overlap between two individuals’ network of friends is associated with the strength of their ties to one another (Granovetter 1973). In that sense, social ties can be classified as either strong or weak. Strong ties constitute stronger and closer relationships that are within an individual’s personal network (Brown and Reingen 1987). People have a wide range of social networks among which to search for information, and this includes both strong ties, such as family members and close friends and weak ties, such as acquaintances. However, dynamic information seeking and product referral are more likely to happen among relationships with strong ties (Brown and Reingen 1987).
Similar to the offline environment, in online settings such as SNS, there exist varying degrees of social relationships among members, and these can also be classified as either strong or weak (Chu and Kim 2011). The perceived tie strength established via SNS motivates consumers to communicate with one another and to disseminate product-related information (Chu and Kim 2011). Although both strong and weak ties contribute to the propagation of eWOM on SNS, weak ties exert a wider impact by extending consumers’ personal network to external communities. On the other hand, strong ties have a more important impact at the individual and small group level; therefore, similar to the offline
20


environment, information-seeking and referral behavior is more likely to happen among those with strong ties (Chu and Kim 2011). In fact, consumers trust online information posted by friends with whom they perceive a strong social relationship more than information posted by those with whom they perceive a weak social relationship (Pan and Chiou 2011). Thus, Facebook members may perceive Facebook as a valuable platform for accessing eWOM because they can access eWOM from close ties. In addition, the emotions triggered by reading Facebook posts from users with strong ties are higher compared to posts from those with weak ties (Lin and Utz 2015). Consistently, it can be expected that Facebook members’ level of excitement derived from checking eWOM posted by close ties is higher compared to weak ties. Therefore, we propose the following:
//j. Tie strength has a positive effect on the cognitive attitude.
H6: Tie strength has a positive effect on affective attitude.
Prior studies on information systems reported that cognitive attitude is the main driver for the adoption of information technology (Angst and Agarwal 2009; Bhattacherjee and Sanford 2006). Cognitive attitude is also examined in eWOM adoption studies of online consumer review platforms. These prior studies conceptualized the cognitive attitude in terms of consumers’ belief about the credibility of eWOM and found that the perceived credibility of eWOM is positively associated with the eWOM adoption. Cognitive attitude is also conceptualized as the degree to which the use of a specific platform for eWOM is perceived to be beneficial by consumers (Leung et al. 2015; Shih et al. 2013). This specific perspective is also relevant to Facebook, where the main goal of the platform is to maintain social relationships among its users. If a Facebook user believes that using Facebook to access eWOM is valuable, he/she is more likely to adopt the eWOM. Therefore, we the following hypothesis is proposed next:
21


H7: Cognitive attitude has a positive effect on the eWOM adoption.
In addition to cognitive attitude, consumers’ adoption of eWOM can be examined through the lens of their affective attitude change (Fang 2014; Xu 2014). The affective attitude might prompt judgmental responses that may be faster and more consistent across individuals (Shih et al. 2013). For instance, studies in marketing and advertising maintain that consumers’ affective attitude toward advertisements affects the consumers’ behavioral intention to adopt the ads (Edell and Burke 1984; Huang et al. 2013; Leung et al. 2015; MacKenzie et al. 1986; Shimp 1981). On Facebook, we can also consider implicit eWOM (i.e., information shared about a product through “Likes” and “Check-ins”) as a source of advertisement for a product; thus, consumers’ adoption of implicit eWOM can be examined through changes in their affective attitude. In addition, a study by Fang (2015) showed that the level of arousal (affective attitude) derived from reading explicit eWOM affects eWOM adoption. Therefore, by considering the notion that Facebook encompasses both explicit and implicit eWOM, we argue that the consumers’ arousal derived through exposure to eWOM on Facebook affects their behavioral intention to adopt the eWOM:
Hs: Affective attitude has a positive effect on the eWOM adoption.
The ELM specifies that consumers use a combination of central and peripheral factors in processing information (Bhattacheij ee and Sanford 2006; Petty and Cacioppo 1986). Prior studies have maintained that central factors (i.e., factors related to the review text) have a stronger influence in changing consumers’ attitude because of the higher amount of cognitive effort required to process information embodied by the review text (Baek et al. 2012; Bhattacheij ee and Sanford 2006). On the other hand, affective attitude change happens quickly and through heuristic factors such as social and emotional responses toward the reviewer (Fang 2014; Xu 2014). Since the eWOM available on Facebook includes a
22


combination of both explicit and implicit eWOM, we believe the eWOM type (explicit vs. implicit) moderates the effect of cognitive and affective attitude on eWOM adoption. The effect of cognitive attitude on eWOM adoption is higher if the eWOM is explicit and the effect of affective attitude on eWOM adoption is higher if eWOM is implicit. Therefore, we propose the following:
H9a'- The effect of cognitive attitude on eWOM adoption is higher for explicit eWOM.
Hn'. The effect of affective attitude on eWOM adoption is higher for implicit eWOM.
Research Methodology
Measurement of Constructs
The scales used to measure the constructs were adopted from the previous literature, with some alterations made to fit the context of this study. The instrument was pilot tested with a sample comprising two professors, two Ph.D. students, and 50 university students with appropriate knowledge and experience of using SNS to assess the face and content validity of the measures. The purpose of the pilot study was threefold: (1) to assess the internal and external validity of the scale items; (2) to estimate potential participation rates for the study; and (3) to provide insight into blind spots and oversights that must be addressed in order to execute the research plan.
The central route of the ELM in our model consists of the product-related attributes of the review (PRA). PRA is defined as the degree to which a review reflects the characteristics of the product or service. The scales for measuring this construct were adapted from (Qiu et al. 2012). The source credibility (SC) represents the peripheral route of the ELM in our model. It is defined as the eWOM recipients’ perception of the sender’s credibility. The scales to measure this construct were adapted from multiple previous studies (Bhattacherjee and Sanford 2006; Cheung et al. 2012; Cheung et al. 2009). The cognitive attitude (COGA)
23


reflects the attitude of a person about using Facebook to access online reviews. The scales used to measure this construct were adapted from (Shih et al. 2013). eWOM adoption (EA) is defined as the degree to which an individual on Facebook perceives that a friend’s post on Facebook about a product/service was informative for the individual’s purchase decisionmaking process. The scales to measure this construct were adapted from (Cheung et al.
2009). Peer image building (IMG) is defined as the degree to which someone perceives that a friend uses Facebook to build his/her social image. The scales for measuring this construct were adapted from (Luam et al. 2015). Tie strength (TS) reflects an individual’s degree of closeness with, and perceived importance of, a Facebook friend. The scales to measure this construct were adapted from (Chu and Kim 2011). Affective attitude (AFFA) reflects the individual’s feelings about sharing products and brand information through Facebook. The scales used to measure this construct were adapted from (Fang 2014). The eWOM type was measured by a binary variable, where “0” represents explicit eWOM and “1” represents implicit eWOM.
Data Collection
We used a survey to collect data from students at two large public universities in the western United States. Prior studies have shown that students are a good representative for empirical studies on SNS and constitute the major percentage of Facebook users (Chu and Kim 2011; Lenhart 2009). Empirical data for testing of the hypotheses were collected from business major undergraduate students in October 2016. We began by providing the students with a brief description of the survey, without revealing our hypotheses. In the first section, we asked questions about the student’s demographic information. Information about their cognitive and affective attitudes was also collected in this section. In the next section, we
24


asked students to recall a friend who had engaged in eWOM communication within the past month, and then we asked the respondents to select one of the following to describe his/her friend’s eWOM activity: (1) he/she posted a review of a product/service on his/her Facebook page; or (2) he/she shared product/service information by “Liking” a Facebook page related to the product/service or by “Checking-in” at the business. Then we asked participants to answer questions pertaining to the product-related attributes of the review, the friend’s credibility, tie strength, and image building, as well as participants’ eWOM adoption.
Sample Profile
In all, a total of 202 usable questionnaires were collected. Forty-five percent (45%) of the respondents were female and 55 percent were male. Most of the respondents (50%) were within the age range of 18-21 years, followed by 30 percent within the ages of 22-25 years, and 20 percent over the age of 25. Most of the respondents in our sample had between 200 and 300 friends on Facebook.
Data Analysis
We used partial least squares (SmartPLS version 3.0) to test the measurement model and the structural model. Partial least squares (PLS) analysis was chosen over other analytical techniques for two reasons. First, it simultaneously tests both the measurement model and the structural model. Second, it is more appropriate for analyzing moderating effects, because traditional techniques cannot account for measurement error in exogenous constructs (Chin 1998a; Chin 1998b; Chin et al. 2003).
Measurement Model Analysis
To examine the psychometric properties of the measurement model, this study examined the composite reliability, convergent validity, and discriminant validity of the
25


constructs. Convergent validity can be assessed by examination of the measurement’s model loadings. These loadings once deemed consistent with the underlying construct, were used to assess internal consistency and average variance extracted (AVE). Convergent and discriminant validity is adequate for constructs modeled using two or more reflective indicators when: (1) all the constructs’ AVE values are above 0.5; and (2) item loadings exceed 0.70 and load more highly on the constructs they are intended to measure (Chin 1998a; Chin 1998b). Table 2 shows the composite reliability, average variance extracted, and Cronbach’s alpha. The composite reliability of the constructs was above the recommended benchmark of 0.7 (Barclay et al. 1995; Chin 1998a; Chin 1998b). All of the constructs’ AVE values were above the recommended level of 0.5 (Chin 1998a; Chin 1998b). Therefore, we found the measurement model’s convergent validity to be satisfactory. Table 3 shows that the square root of the AVE value for each construct exceeds the correlation between that construct and other constructs (Chin 1998a; Chin 1998b; Fornell and Larcker 1981), thus providing evidence for discriminant validity.
Common Method Bias
Because survey methodologies may be subject to common method bias (CMB), we ran a PLS test for CMB using the common factor approach described by Liang et al. (2007). We created a common method construct having all the items associated with it; we then modeled each of the 30 indicators as a single-indicator construct and created paths between them and the common method construct as well as the theoretical constructs. The results showed that loadings on the theoretical constructs were both high and highly significant. Loadings on the common method construct were low and, in almost all cases, nonsignificant. This indicates that CMB is not a problem in this research (Liang et al. 2007).
26


Structural Model
Unlike covariance-based SEM, PLS does not provide summary statistics to allow for assessment of the overall “fit” of the model. However, the variance explained by the path model (R-squared) of the endogenous construct—the eWOM adoption, and the sign and significance of the path coefficients—is typically used to assess model fit. A bootstrapping approach was used to produce 500 random samples of the original sample size from the data set by sampling through replacement. This was necessary to obtain estimates of the standard errors for use in testing the statistical significance of the path coefficients. Such an approach provides valid estimates of the significance of the path coefficients in the PLS models (Mooney et al. 1993). A summary of the empirical testing and validation of the theorized casual links is given in Figure 2. The product-related attributes of the review, source credibility, peer image building, and tie strength were found to explain 42.6 percent of the variance in cognitive attitude. The image building and tie strength explained 14.9 percent of the variance in affective attitude. The cognitive attitude and affective attitude, in turn, explained 36.9 percent of the variance in eWOM adoption. As we predicted, the effect of the product-related attributes of the review on cognitive attitude was significant (/?=0.167, P<0.05), supporting Hi. The effect of source credibility on cognitive attitude was also significant (/?=0.483, P<0.01), supporting H2. We found that peer image building had a negative effect on the cognitive attitude (/?= -0.186, P<0.01) and affective attitude (fi= -0.342, P<0.01), supporting H3 and H4, respectively. The effect of tie strength on cognitive attitude (H5) was not significant; however, the effect of tie strength on the affective attitude was significant (/?= 0.178, P<0.05), supporting Hi. As we predicted, the effect of cognitive attitude on the eWOM adoption was significant (/5=0.334, P<0.01), supporting H7. The effect
27


of affective attitude on the eWOM adoption too was significant (/?=0.254, P<0.01), supporting Hx. Consistent with H9a and H9b, the eWOM type was found to moderate the effect of cognitive attitude and affective attitude on the eWOM adoption. We also found that, for explicit eWOM, cognitive attitude had a higher effect on the eWOM adoption (fi= -0.30, P<0.01) and, for implicit eWOM, affective attitude had a higher effect on the eWOM adoption (J3= 0.353, P<0.01).
Discussion
The main purpose of this study is to move beyond the conventional definition of eWOM as a form of reviews delivered through written text (i.e., explicit eWOM) by highlighting the notion that non-textual information such as “Likes” and “Check-ins” on Facebook can be considered as a new type of eWOM (i.e., implicit eWOM). Against this background, we consider Facebook as an eWOM communication platform that encompasses both explicit and implicit eWOM, and by building on the ELM and the affect-as-information theory, we develop a unified theoretical model that examines the role of eWOM type in the eWOM adoption process on Facebook.
The findings of this study are threefold. First, as with studies of eWOM adoption in online consumer review platforms and based on the theoretical lens of the ELM, we find support for the positive effect of cognitive attitude on eWOM adoption. In addition, based on the affect-as-information theory, we examined the explanatory power of the affective path in eWOM adoption by finding support for the positive effect of affective attitude on eWOM adoption. Reflecting on affective attitude, this study confirms the findings of Fang (2015) concerning the effects of the arousal dimension of affective attitude on eWOM adoption, and it extends the notion of arousal to the emotional outcome of exposure to both explicit and
28


implicit eWOM. Second, regarding the role of eWOM type in the eWOM adoption process, we find that eWOM type moderates the effect of cognitive and affective attitudes on the eWOM adoption. Our results show that, when the eWOM is explicit, as it is in online consumer review platforms, the cognitive attitude has a higher effect on the eWOM adoption. However, if the eWOM type is implicit, the affective attitude has a higher effect on the eWOM adoption. Third, regarding the drivers of cognitive and affective attitude, our results confirm the importance of review text (i.e., the ELM’s central cue) as one of the antecedents of cognitive attitude. Our results show that, when it comes to explicit eWOM, Facebook users pay attention to product-related attributes of the explicit eWOM, and the higher elaboration of those attributes resulted in a higher cognitive attitude. In addition, our results support prior eWOM adoption studies that highlight source credibility as the most important peripheral factor for explicit eWOM adoption. Our model considered the socio-interpersonal relationships among Facebook members and examined the effects of tie strength and peer image on the users’ cognitive and affective attitudes. We did not find support for a positive effect of tie strength on the cognitive attitude. One possible explanation for this result might be the finding from prior studies that consumers accept eWOM from their close ties because they perceive them as a credible source of eWOM. Thus, tie strength may have an indirect effect on cognitive attitude by influencing perceived source credibility. We further checked this path in our structural model and found that tie strength had a positive effect on the source credibility (/?= 0.44, P<0.01). However, we also found a positive effect of tie strength on the affective attitude, providing support for the role of tie strength as one of the drivers of affective attitude. Our results also highlight the role of image building as one of less-studied personal conditions that motivates eWOM behavior in SNS, by providing empirical support
29


for image building having a negative effect on eWOM adoption by having a negative influence on both the cognitive and affective attitudes.
Implications
Theoretical Implication
This study makes important theoretical contributions to the extant body of literature on eWOM communication. The question of how and under what circumstances consumers adopt eWOM has become an important research topic. Most previous eWOM studies have examined the eWOM adoption from online consumer review sites (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2008; Cheung et al. 2009; Mudambi and Schuff 2010). These studies considered the perceived credibility of online reviews as the most important driver of eWOM adoption (Baek et al. 2012; Wang 2010). Other studies, in a similar vein, collected online reviews to examine how the attributes of a review’s text and the characteristics of the eWOM source influence the perceived helpfulness of online reviews, as the flip side of review credibility (Cheung et al. 2012; Salehan et al. 2015; Yin et al. 2014a). This study contributes to this body of literature by examining the eWOM adoption in social networking services, and in particular, Facebook. Despite the emergence of SNS as a new platform for eWOM communication, the eWOM adoption on SNS has been largely neglected in previous studies. In particular, the eWOM adoption on Facebook is important because of the differences and commonalities between Facebook and the conventional platforms for eWOM communication. These differences stem from the notion that social presence and interpersonal relationships between members are the primary reasons for the use of Facebook (Fang 2014; Kaplan and Haenlein 2010); in other words, the communication of eWOM itself is not the main emphasis of Facebook. A second important distinction is that, in addition to
30


explicit eWOM, there is evidence that non-textual signals such as information shared through users’ “Likes” and “Check-ins” have the potential to influence purchase decision making of Facebook users (Keenan and Shiri 2009; Luarn et al. 2015; Okazaki 2009). We refer to these non-textual forms of eWOM here as implicit eWOM. Reflecting on the contextual differences between Facebook and conventional eWOM communication platforms, and by considering eWOM type, this study contributes to the theoretical underpinning on eWOM adoption studies in the following ways.
First, this study contributes to the conceptualization of cognitive and affective attitudes as the two antecedents of eWOM adoption. Prior eWOM adoption studies of both online consumer review platforms (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009) and Facebook (Fang 2014) have examined the perceived credibility of the eWOM as the focal cognitive-based construct in their theoretical model (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009). However, we find that individuals’ perception of the suitability of Facebook for accessing online reviews is an important cognitive-based construct for examining the adoption of eWOM on Facebook. Thus, our findings confirm prior studies (Leung et al. 2015; Shih et al. 2013) in taking into account the usability of the platform for eWOM communications when the social relationships among members of the platform is an integral part of the platform. In addition, while the role of affect in consumers’ decision making has been examined in past studies (Eroglu et al. 2003; Forgas and George 2001; Soderlund and Rosengren 2007), few studies have examined the affective attitude as one of the antecedents of eWOM behavior on Facebook (Fang 2014; Shih et al. 2013; Xu 2014). This study contributes to this stream of research by finding evidence that the arousal dimension of the affective attitude affects eWOM adoption. In particular, this study extends
31


the findings from Fang (2015) on the effect of arousal on eWOM adoption by providing evidence that the arousal derived from exposure to both explicit and implicit eWOM can serve as a source of information and thus can affect eWOM adoption.
Second, this study confirms the suitability of the ELM and the affect-as-information theory to understand the drivers of the cognitive and affective attitude, thus providing empirical support for the integration of these two theories to understand eWOM adoption on Facebook. Similar to previous studies of online consumer review platforms (Ghose and Ipeirotis 2011; Qiu et al. 2012; Schindler and Bickart 2012), our results show that product-related attributes of a review on Facebook have a positive effect on the cognitive attitude.
Our results also confirm prior studies that identify source credibility as the most important of ELM’s peripheral factors affecting cognitive attitude (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009). In addition, this study examined tie strength and peer image as two social-interpersonal peripheral factors that affect both cognitive and affective attitudes. While prior studies examined tie strength as one of drivers of eWOM engagement as well as information seeking and giving behavior in SNS (Chu and Kim 2011; Hansen and Lee 2013; Pan and Chiou 2011), our results show that tie strength did not have a direct effect on the cognitive attitude. However, further analysis showed that it had an indirect effect on cognitive attitude by having a positive effect on perceived source credibility. Our results also show that peer image building has a negative effect on cognitive attitude, thus contributing to prior literature (Folkes 1988; Sen and Lerman 2007) that examined the role of internal (self-serving) and external (product-related) motives in adopting product recommendations. Our results show that image building behavior among Facebook users can be considered as an internal motive by Facebook members and that Facebook users tend to discount eWOM
32


generated by users who use published content such as product and brand information to promote their social image.
Third, this study examined the role of eWOM type in the eWOM adoption process by finding that the eWOM type moderates the effect of the cognitive and affective attitudes on the eWOM adoption process. Our results show that the cognitive attitude has a higher effect on eWOM adoption when the eWOM is explicit. Conversely, the affective attitude has a higher effect on eWOM adoption when the eWOM is implicit.
Practical Implication
The results of this study also offer two important practical contributions. First, our results show that, as with online consumer review platforms, the content of Facebook reviews plays an important role in consumers’ decision-making process. Given Facebook’s role as a platform where the main purpose is hedonic rather than for communication of eWOM, the perception of Facebook users about Facebook’s suitability for accessing online reviews is particularly important for eWOM adoption from this platform. Our results show that, when it comes to explicit eWOM, Facebook does not differ significantly from online consumer review platforms, and the extent to which the eWOM elaborates on the attributes of a product positively influences Facebook users’ perception of the suitability of Facebook for accessing online reviews. This finding may hold important implications for the current practice of social advertisement in online shopping platforms. Many online shopping platforms allow buyers to share their purchase information on an SNS immediately after their purchase. For instance, Figure 31 shows an example of an Amazon purchase, where the buyer can share information about his/her purchase through various channels, including Facebook.
1 Source: https://vwo.com/blog/high-converting-thank-you-pages/
33


The content and metadata for this purchase are pre-filled with generic content. As our result shows that the product-related attributes of the eWOM are positively associated with eWOM adoption, our suggestion for online retailers and shopping platforms would be to persuade Facebook users to modify the content of such purchase information based on the attributes of the product that they like best.
Second, our results show that the affective attitude of consumers toward eWOM on Facebook is one of the drivers of eWOM adoption on Facebook and that the affective attitude has a higher effect on eWOM adoption when the eWOM type is implicit. In addition, our results show that tie strength and peer image are two social-interpersonal relationship variables that influence affective attitude. These findings have important implications for the practice of sponsored story advertising on Facebook. Such sponsored ads are commonly accompanied by social cues such as businesses/products or brands that other Facebook friends have either “Checked-in” or “Liked.” A prior study using a randomized field experiment showed that the presence of close ties as a social signal increased the likelihood of response to such advertisements (Bakshy et al. 2012). While our study confirms this finding, the notion of image building behavior among Facebook users demands further attention for the practice of sponsored advertisement. While further investigation would be needed in order to quantify the measure of image building, we believe that one possible approach to mitigate the negative effect of image building on eWOM adoption is to consider Facebook users’ past “Likes,” “Check-ins,” and other socio-economic signals before including them as a social component of a sponsored advertisement. For instance, if a Facebook user’s new brand page “Like” deviates from his/her past “Likes” concerning the same product or service type, this new “Like” can be excluded from the sponsored
34


advertisement. Alternatively, if multiple Facebook friends have “Liked” the same product/brand page, because of the positive effect of a number of social cues on the likelihood of response to an advertisement (Bakshy et al. 2012), these Facebook members can be included in the advertisement but not as the main social cue.
Limitations and Future Research Direction One of the major limitations of this study is the use of university students as the sample respondents, which may raise questions about the generalizability of the results. Although previous studies have shown that students are a good representative of the population for empirical studies on SNS, and that they are known to constitute the major percentage of Facebook users (Chu and Kim 2011; Lenhart 2009), our findings should be generalized only with caution. Our sample also lacks cultural and language diversity, which may limit its general application to other cultures. Language, as the vehicle for a textual message, has a significant effect on the perception of the receiver; hence, the adoption of eWOM may be different across diverse languages. Also, a study by(Chu and Choi 2011) showed that differences in culture influence the acceptance of eWOM. People from collectivist cultures engage in greater levels of information-seeking and information-giving behaviors on SNS than their individualistic counterparts. Moreover, there are differences between collectivist and individualistic societies in terms of network structure. People from collectivist cultures have more strong ties on SNS, while people from individualistic cultures tend to have weaker ties. Finally, people from collectivist cultures have higher levels of trust in their SNS contacts, and they are thus more likely to be influenced by eWOM than their individualistic counterparts. Hence, in the future, comparative studies on other social networking services in different cultures would be fruitful. Furthermore, we use self-reported
35


data in this study, which is subject to social desirability response bias (Arnold and Feldman 1981). To address this bias, it may be useful for future research to analyze textual information collected from SNS and to look at how different characteristics of the text, such as readability and descriptiveness, influence its acceptance. Future research may also conduct experiments to investigate the acceptance of eWOM. Previous research has shown that computer-mediated communications (CMC) can effectively transfer emotions. Moreover, the emotions contained in a message transferred through CMC significantly influence how the message is processed and interpreted by its recipient (Berger and Milkman 2012; Riordan and Kreuz 2010). Hence, future research may consider the effect of message sentiment on the acceptance of explicit eWOM. In addition, because of the negative impact of an individual’s image building behavior on eWOM adoption, quantifying this construct based on individuals’ past social behavior would be fruitful for social media marketers.
Conclusion
Although online consumer review websites are the major platform for eWOM communication, the emergence of SNS has created a new avenue for both consumers and marketers to review and share brand/product-related information. To that end, this study highlights the differences between SNS and conventional platforms for eWOM communication. We argue that, while explicit eWOM is common across all platforms for eWOM communication, implicit eWOM is only salient on SNS. This study draws on the ELM and the affect-as-information theory to understand the drivers of cognitive and affective attitudes, thus contributing to prior eWOM adoption studies by providing empirical evidence of the benefits of integrating these two theories to explore the eWOM adoption process in the presence of both explicit and implicit eWOM. Our results show that product-related
36


attributes of a review, positively affect the cognitive attitude. In addition, we find that source credibility also has a positive effect on the cognitive attitude. Furthermore, our study considers tie strength and peer image as two social-interpersonal relationship variables that influence both cognitive and affective attitude. Our results show that image building has a negative effect on both the cognitive and affective attitudes and that tie strength influences eWOM adoption by having a positive effect on the affective attitude; however, we did not find support for the positive effect of tie strength on the cognitive attitude. Our study shows that the eWOM type moderates the effect of cognitive and affective attitudes on the adoption of eWOM. The effect of cognitive attitude on the eWOM adoption is higher when the eWOM is explicit. Conversely, the effect of affective attitude on the eWOM adoption is higher when eWOM is implicit. The results of this study contribute to the extant body of eWOM adoption literature by including eWOM type in our theoretical model of eWOM adoption and by investigating its moderating impact on the effects of cognitive and affective attitudes on eWOM adoption.
37


TABLES AND FIGURES
Figure II-1. Research Model
Table II-1. Measurement Items
Construct Items Reference
Product-related attributions of the review (PRA) PRAi.The content of his/her review was based on the product/service PRA2. His/her review reflects the characteristics of the product/service (Qiu et al. 2012)
Source credibility (SC) SCi. He/she was knowledgeable about the product/service SC2. He/she was trustworthy SC3. He/she was credible SC4. He/she appears to be an expert on the product/service (Bhattacheij ee and Sanford 2006; Cheung et al. 2012; Cheung et al. 2009)
Cognitive attitude (COGA) COGAi. Using Facebook to access reviews about products/services is wise COGA2. Using Facebook to access reviews about products/services is beneficial COGA3. Using Facebook to access reviews about products/services is valuable (Shih et al. 2013)
eWOM adoption (EEA) EAi. The information shared through my friend’s post on Facebook contributed to my knowledge of the product/service (Cheung et al. 2009)
38


Construct Items Reference
EA2. My friend’s post on Facebook about the product/service made it easier for me to make a purchase decision (e.g., to purchase or not to purchase) EA3. My friend’s post on Facebook about the product/service has enhanced my effectiveness in making a purchase decision EA4. My friend’s post on Facebook about the product/service motivated me to make a purchase decision.
Peer image building (PIMG) PIMGi. He/she uses Facebook to shape an impression of himself/herself PIMG2. He/she uses Facebook to build image of himself/herself (Luarn et al. 2015)
Tie strength (TS) TSi. He/she is important to me TS2. He/she is close to me TS3.1 contact him/her frequently (Chu and Kim 2011)
Affective attitude (AFFA) When I see friends share information about products/services on Facebook: AFFAi. I’m excited AFFA2. I’m frenzied AFFA3. I’m wide awake (Fang 2014)
Table 11-2. Reliability and Convergent Validity
Average Variance Extracted (AVE) Composite Reliability Cronbach's Alpha
EA 0.712 0.907 0.862
AFFA 0.829 0.936 0.897
COGA 0.932 0.976 0.964
PIMG 0.94 0.969 0.936
PRA 0.874 0.932 0.862
SC 0.802 0.942 0.917
TS 0.888 0.96 0.937
39


Table II-3. Discriminant Validity
EA AFFA COGA PIMG PRA SC TS
EA 0.844
AFFA 0.284 0.911
COGA 0.388 0.403 0.966
PIMG -0.329 -0.343 -0.344 0.969
PRA 0.591 0.162 0.46 -0.197 0.935
SC 0.527 0.346 0.608 -0.258 0.554 0.896
TS 0.346 0.18 0.234 -0.008 0.333 0.44 0.942
Note: Diagonal values are square roots of A
fEs.
Table II-4. Cross Loadings
Affective Cognitive Image Tie Adoption SC PRA
AATTi 0.902 0.372 -0.325 0.17 0.303 0.339 0.173
aatt2 0.927 0.357 -0.326 0.172 0.227 0.29 0.096
aatt3 0.903 0.372 -0.283 0.149 0.241 0.312 0.173
CATTi 0.408 0.959 -0.328 0.232 0.344 0.584 0.417
catt2 0.38 0.968 -0.321 0.229 0.362 0.588 0.453
catt3 0.381 0.969 -0.346 0.218 0.415 0.589 0.461
PIMGi -0.346 -0.339 0.971 -0.027 -0.312 -0.246 -0.213
pimg2 -0.318 -0.327 0.968 0.012 -0.328 -0.256 -0.168
TSi 0.177 0.238 -0.004 0.941 0.269 0.428 0.333
ts2 0.213 0.219 -0.018 0.966 0.342 0.425 0.297
ts3 0.104 0.201 0 0.92 0.382 0.385 0.313
EAi 0.227 0.355 -0.36 0.253 0.81 0.471 0.669
ea2 0.242 0.419 -0.341 0.31 0.918 0.525 0.583
ea3 0.269 0.335 -0.257 0.343 0.91 0.482 0.467
ea4 0.224 0.16 -0.122 0.258 0.721 0.263 0.226
SCi 0.292 0.607 -0.252 0.382 0.526 0.892 0.54
sc2 0.321 0.537 -0.205 0.45 0.45 0.931 0.51
sc3 0.296 0.522 -0.212 0.414 0.441 0.92 0.498
sc4 0.332 0.501 -0.254 0.328 0.463 0.836 0.428
PRAi 0.118 0.323 -0.175 0.276 0.507 0.445 0.906
pra2 0.174 0.503 -0.192 0.337 0.587 0.57 0.962
40


Table II-5. Common Method Bias
Indicators Theoretical Construct Loading T-stat P-value Common Method Factor Loading T-stat P-value
AATTi 0.861 25.553 p<0.01 0.053 1.158 p>0.05
aatt2 0.954 46.329 p<0.01 -0.043 1.314 p>0.05
aatt3 0.917 36.941 p<0.01 -0.008 0.191 p>0.05
CATTi 0.973 45.263 p<0.01 -0.017 0.707 p>0.05
CATT2 0.976 54.003 p<0.01 -0.011 0.425 p>0.05
CATT3 0.948 50.232 p<0.01 0.027 1.206 p>0.05
EAi 0.681 8.5 p<0.01 0.161 1.876 p>0.05
ea2 0.828 16.811 p<0.01 0.107 1.825 p>0.05
ea3 0.938 28.29 p<0.01 -0.028 0.618 p>0.05
EAi 0.939 10.789 p<0.01 -0.272 2.782 p<0.01
PIMGi 0.966 90.527 p<0.01 -0.008 0.456 p>0.05
pimg2 0.973 96.396 p<0.01 0.008 0.456 p>0.05
PRAi 1.025 41.234 p<0.01 -0.126 3.431 p<0.01
pra2 0.85 23.153 p<0.01 0.126 3.431 p<0.01
SCi 0.736 11.682 p<0.01 0.168 2.628 p<0.01
sc2 0.992 19.627 p<0.01 -0.064 1.101 p>0.05
sc3 1.024 19.192 p<0.01 -0.112 1.686 p>0.05
sc4 0.818 9.969 p<0.01 0.019 0.202 p>0.05
TSi 0.93 39.519 p<0.01 0.005 0.149 p>0.05
ts2 0.96 56.015 p<0.01 0.007 0.23 p>0.05
ts3 0.939 43.716 p<0.01 -0.012 0.322 p>0.05
41


Figure II-2. Empirical Results
m
y3 f acebook Twitter Is—1 E-mail
^jiar^Jh^Ogen^gglei^an^i^nemies^oljyh^
1
^^ell of Plato
I just bought ’The Open Society and Its Enemies. Vol. 1: The Spell of Plato* by Kan R
Pocoer was born *â– > 1902 to a Viennese famly of Jewish origin He taught m Austria unti 1937. w
j Sfiare Uiis item J
olowng year, arx Your shann
Figure II-3. Social Advertising in Amazon
42


CHAPTER III.
THE EFFECT OF REVIEW CONSISTENCY ON USEFULNESS OF ONLINE REVIEWS: EVIDENCE FROM REVIEWS IN SERVICE CONTEXT
Abstract
Online reviews play a significant role in product diagnostics as well as in the formation of consumers’ purchase intention. The integration of consumer review features in online retail websites such as amazon.com and online consumer review websites such as yelp.com has enabled users to share their experiences with other consumers seeking trustworthy information about specific products and services. Further, the social voting features in Yelp and Amazon are examples of mechanisms that augment trust by allowing consumers to vote for the usefulness of online reviews. Employing the Elaboration Likelihood Model (ELM), this study analyzes Yelp data to better understand the antecedents of review usefulness. The data analyses show that unsupervised and supervised measures of review consistency, representing both the central and peripheral cues of ELM, have positive effects on review usefulness. The analyses also show the negative effect of review rating and the positive effect of source credibility, reflecting the non-trivial effect of the peripheral cues of ELM. These findings provide new insights on how sentiment-mining and machinelearning techniques can be utilized to analyze the quality of online reviews. Our findings also provide insights for online retailers on how to manage and sort reviews in their websites. Keywords: online review, review consistency, Yelp, social voting, e-Word of Mouth, machine learning, text mining, sentiment analysis.
43


Introduction
Advances in information technology and the emergence of Web 2.0 have enabled consumers to freely share their opinions and experiences about products and services anytime, anywhere. Like other forms of electronic word of mouth (eWOM), online reviews play an inescapable role in reducing the uncertainty in online shopping by guiding consumers in their purchase decision making process (Kuan et al. 2015; Yin et al. 2014a). It has been reported that 92 percent of consumers read online reviews prior to purchases and 89 percent of consumers make decision in their purchase based on online reviews (eMarketer 2010). Online retailer websites such as Amazon and online consumer review websites such as Yelp are among the most used product recommendation platforms for online consumers.
According to a recent report (Nielsen 2012), after family members and friends, consumer review websites are rated as the most trustworthy and influential source of information on products and services. This means that people view online reviews as more reliable and less biased than any other information available on products and services (Lee and Youn 2009).
However, the overwhelming volume of online reviews presents challenges for consumers seeking relevant and trustworthy reviews (Baek et al. 2012; Kuan et al. 2015; Mudambi and Schuff 2010). To address this problem, online consumer review websites such as Yelp and Amazon have implemented online social voting systems. An online social voting system is a technological component that allows consumers to diagnose the credibility of online reviews by voting for its usefulness or helpfulness (Baek et al. 2012). Understanding the factors that impact the helpfulness/usefulness of online reviews is an important question for both academics and practitioners because online review quality is directly related to consumers’ level of trust in the online marketplace. In addition, from a theoretical
44


perspective, review usefulness (helpfulness), as the proxy for review credibility, is known to be the most important driver for consumers to adopt eWOM (electronic word-of-mouth) (Cheung et al. 2012; Cheung et al. 2009). Other studies have reported that online reviews that are perceived as useful exert a stronger influence on an online user’s decision to conduct online transactions than less useful reviews (Baek et al. 2012; Yin et al. 2014a). From a practical perspective, online retailers that display helpful reviews gain a strategic advantage in consumer attention and “stickiness” (Connors et al. 2011; Yin et al. 2014a). Reflecting the important role of social voting systems in the online marketplace, Amazon added an extra $2.7 billion to its revenue by implementing its social voting system and asking consumers to rate the helpfulness of its online reviews (Spool 2009).
The extant body of research on online reviews has mainly focused on identifying the factors that affect the usefulness (in Yelp) or helpfulness (in Amazon) of online reviews (Baek et al. 2012; Cao et al. 2011; Forman et al. 2008; Kuan et al. 2015; Mudambi and Schuff 2010; Salehan and Kim 2014; Salehan and Kim 2016; Zhang et al. 2010). They have examined the relationship between characteristics of online reviewers and review quality, or the relationship between the number of star rating and review quality, where the review quality is operationalized as the number of useful votes (in Yelp) or helpful votes (in Amazon) received from other online users (Forman et al. 2008; Mudambi and Schuff 2010). More recent studies have begun to use text mining and sentiment mining methods to investigate how the online review content impacts review usefulness (Baek et al. 2012; Kuan et al. 2015; Salehan et al. 2015). These studies focus on the effect of the valence of online reviews (i.e., negative, neutral, and positive) or the emotion-bearing words in the review on review helpfulness.
45


These prior studies focus either on non-verbal features of review rating or verbal feature of review text, failing to consider that both review text and review rating are inseparably entangled signals of reviewers’ attitude about products/services (Tang and Guo 2015) . In practice, consumers tend to simultaneously consider both non-verbal star rating and verbal review text to assess the usefulness of an online review, therefore, inconsistency between a review text and the accompanying numerical rating increases the uncertainty and decreases the value of online reviews (Mudambi et al. 2014).In addition, prior studies reported the biases associated with using single-factor measure (i.e. numerical star rating) in evaluating online reviews. For instance, consider the below text about Sinan Aral’s2 experience on Yelp:
“A few months ago, I stopped in for a quick bite to eat at Dojo, a restaurant in New York City’s Greenwich Village. I had an idea of what I thought of the place. Of course, I did — I ate there and experienced it for myself. The food was okay. The service was okay. On average, it was average. So I went to rate the restaurant on Yelp with a strong idea of the star rating I would give it. I logged in, navigated to the page and clicked the button to write the review. I saw that, immediately to the right of where I would “click to rate,” a Yelp user named Shar H. was waxing poetic about Dojo’s “fresh and amazing, sweet and tart ginger dressing” — right under her bright red five-star rating. I couldn’t help but be moved. I had thought the place deserved a three, but Shar had a point: As she put it, “the prices here are amazing!” Her review moved me. And I gave the place a four. As it turns out, my behavior is not uncommon. In fact, this type of social influence is dramatically biasing online ratings — one of the most trusted sources of consumer confidence in e-commerce decisions.”
As it can be seen, Sinan Aral attributes the bias in numerical star rating to positive
social influence that exists in e-commerce websites. Other studies also maintain that a single
factor measure (i.e., star rating) is not a sufficient proxy for review quality. For example, the
J-shaped distribution of star rating, the phenomena that “most products reviewed receive
almost five stars” and very few products receive one or two stars, has led researchers to
2 http://sloanreview.mit.edu/article/the-problem-with-online-ratings-2/
46


question the reliability of online review system itself (Gao et al. 2015a; Hu et al. 2009). Hu et al. (2009) point out that people tend to post online reviews when they are extremely happy or extremely unhappy. That means people who feel an average level of satisfaction with a purchased product or service might not bother to post a review. This suggests that “people should not solely rely on the simple average [star rating] that is readily available,” but instead should incorporate other aspects to make better decisions about the usefulness of a specific online review (Hu et al. 2009).
Reflecting on these prior studies, we theorize the consistency of online reviews, meaning the level of consistency between non-verbal star rating and verbal review text. By drawing upon Elaboration Likelihood Model (Petty and Cacioppo 1986), we explore the factors that affect the review quality, which is operationalized as the number of “useful” votes received by online reviewers. In particular, to address the inconsistency problem in online reviews, we craft and operationalize a review consistency variable, which is conceptualized as a two-factor measure and operationalized as the degree of consistency between the review text (i.e., the qualitative assessment of the online review text, as a central cue) and the numerical star rating (i.e., the quantitative rating of the online review, as a peripheral cue). By using the two-factor measure (i.e., the review consistency variable) and other theoretically important variables, we explore the factors that are important in identifying useful and trustworthy online reviews.
For this exploration, we collect and analyze 2,634 restaurant reviews from Yelp. We use Yelp as the context of this study because consumers rely heavily on online reviews about services due to higher uncertainty associated with services compared to products (Racherla and Friske 2012). In addition, prior research shows that the inconsistency in online reviews is
47


more prevalent in experience goods (e.g., food experiences in restaurants), compared to search goods (Mudambi et al. 2014). To empirically test our proposed model (refer to Figure 3), we employ sentiment-mining and machine-learning techniques in combination with ordered logit and negative binomial regression. The data analyses show that unsupervised and supervised measures of review consistency have positive effects on review usefulness. In addition, the new two-factor variable of review consistency, which we developed for this study, predicts the usefulness of online reviews above and beyond prior single-factor variables.
This study makes important contributions to both theory and practice. In theory, it advances our knowledge of the factors that are important to identify useful and trustworthy online reviews. In particular, the review consistency variable significantly advances ELM by presenting empirical evidence that an online review system becomes more useful3 and trustworthy when the non-verbal star rating (i.e., a peripheral cue) is complemented with the verbal content of online review (i.e., a central cue). In that regard, this study also contributes to the existing literature on online trust but in the new context of online reviews. In practice, this study makes an important contribution, as online review platform providers may improve their sorting algorithms by incorporating the proposed review consistency variable into their recommendation systems.
The remainder of this paper is structured as follows. The next section discusses the suitability of ELM as the theoretical premise of this study. The research model and related hypotheses are presented in this section. In the research methodology section, we discuss the
3 Since review usefulness and review helpfulness are the same mechanism implemented in social voting system of Yelp and Amazon respectively, from now on in this study we just use the term usefulness for the sake of consistency and avoiding any confusions.
48


data, measurements, and econometric model used in this study. Methods of developing measurements for review consistency are also discussed in detail. In the discussion section, the results of our data analysis are reported. Finally, theoretical and practical implications of this study are discussed, followed by a treatment of the limitations of this research and future directions for study.
Theoretical Foundation
Elaboration Likelihood Model
According to Elaboration Likelihood Model (ELM) (Petty and Cacioppo 1986), the persuasiveness of a message can be determined by the central and peripheral cues. The central cues require an individual to think critically about the issues involved in the argument of the verbal message and to inspect the relative facts and relevance of the message before making a conversant judgment about the message. Therefore, processing a message through the central cues demands high cognitive effort. Conversely, processing a message through the peripheral cues involves less cognitive effort, because the influence of a message is highly associated with heuristic cues (Angst and Agarwal 2009). ELM posits that central and peripheral cues affect the judgment of the message recipients to be either persuaded or not persuaded by the received message. The central cues have a persistent effect on the receiver’s attitude, which in turn, influences the receiver’s judgment regarding the message. On the other hand, the peripheral cues have a temporal effect on the attitude of a receiver of the message, which may not lead to the final decision to accept or reject a message. The degree to which central and peripheral cues affect the attitude of an individual depends on the level of motivation and/or message processing capability of the message recipient. If an individual is motivated and/or able to process a message, the central cues are more influential in
49


changing his/her attitude, and in turn, affecting his/her judgment. On the other hand, if an individual is less motivated and/or able to process a message, the peripheral cues are more influential in changing his/her attitude, and thus leading him/her to accept or reject the message (Angst and Agarwal 2009; Bhattacherjee and Sanford 2006; Petty and Cacioppo 1986).
ELM and Review Usefulness
As a technological component of online review systems, the social voting feature, the main focus of this study, allows online users to cast “Useful” votes with respect to online reviews. In particular, the social voting feature in Yelp can be seen as a technological component allowing users to express their views on the credibility of the online reviews themselves (Baek et al. 2012). Review credibility, as the flip side of the perceived usefulness of online reviews, is among the most important concerns in online review research, revealing that reviews with higher credibility exert a greater influence in the adoption of eWOM (Baek et al. 2012; Cheung et al., 2009; Cheung et al., 2012). Prior studies examine perceived credibility of online reviews as a cognitive based attitude construct (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009; Fang 2014). These studies employed ELM’s central and peripheral cues to investigate the antecedents of the review credibility. Consistent with these prior studies, and by considering the notion that online consumer reviews contain both peripheral and central cues (Mousavizadeh et al. 2015), we believe ELM provides a suitable lens for investigating the drivers of review usefulness (as a notion of review credibility) in our study.
Moreover, based on our literature review (refer to Table 1 in the Appendix section), prior studies investigate how the factors related to review text, as well as the
50


attributes of reviewers, affect review usefulness. One of the major limitations of these studies is neglecting how the inconsistency between the central and peripheral cues of ELM affects the usefulness of online reviews. This notion is important from a theoretical perspective as ELM posits that persuasion can be simultaneously executed through the central and peripheral cues (Angst and Agarwal 2009; Bhattacheijee and Sanford 2006; Petty and Cacioppo 1986). To understand this phenomenon, in our research model, we theorize review consistency as a two-factor measure, which reflects the consistency between the central cue (i.e., review text) and the peripheral cue (i.e., numerical star rating), we also consider review sentiment polarity (ELM’s central cue), review rating and reviewer credibility (ELM’ peripheral cues) as theoretically important variables in our research model.
ELM’s Central Cue Review Sentiment Polarity
Past studies have examined the effect of review text as the central cue of ELM on review credibility (Cheung et al. 2012; Cheung et al. 2009). Recent studies examining the content of online reviews find that the emotions embodied in the review text can substantially influence the way a review is processed (Salehan and Kim 2014; Salehan and Kim 2016; Yin et al. 2014b). These findings are consistent with the idea that emotions expressed in a review text perform as a source of social information (Van Kleef 2010). In addition, a recent study by Yin et al. (2014) shows that negative emotions such as anger and anxiety positively influence the usefulness of online reviews. Past studies attribute the positive impact of a negative review on a review’s usefulness to the observed negativity bias in online reviews (Cao et al. 2011; Kuan et al. 2015). This negativity bias implies that negative information is perceived as more informative and diagnostic than positive information (Baumeister et al.
51


2001; Herr et al. 1991). Additionally, negative reviews are more vivid than positive ones, and in an online environment where positive reviews are salient, negative reviews are more likely to gain the attention of consumers (Herr et al. 1991; Kuan et al. 2015). Moreover, negative information reflects a low quality of products, however, positive information can be attributed to both low- and high-quality products (Skowronski and Carlston 1987). Conversely, studies that find that positive ratings are more useful (Carlson and Guha 2010; Chevalier and Mayzlin 2006) argue that, once consumers are predisposed to a product prior to a purchase, they ignore negative ratings because such ratings do not comply with the consumers’ prior belief.
Synthesizing the prior research, we believe that, in the context of a service industry, negativity bias is a better explanation for how a review with negative sentiment polarity affects review usefulness. This is because consumers’ service experiences are inherently different from product experiences. That means, service experiences tend to involve a high level of subjectivity and a high volume of online reviews, leading consumers to find negative information more diagnostic and vivid, thus resulting in negative reviews receiving more useful votes. Therefore, we suggest following hypothesis:
Hi: Reviews with negative sentiment polarity are more likely to receive useful votes
than positive ones.
ELM’s Peripheral Cues
Review Rating
The numerical star rating of an online review is a peripheral factor that reflects a consumer’s view of a product or service quality (Mudambi and Schuff 2010). A one or two-star rating indicates a negative view of a product or service, a three-star rating reflects a moderate view, and a four or five-star rating reflects a positive view of a product or service
52


(Mudambi and Schuff 2010). Consistent with negativity bias, prior studies report a positive
effect of negative ratings on review usefulness (Cao et al. 2011; Kuan et al. 2015; Sen and
Lerman 2007; Willemsen et al. 2011). In line with these prior studies, we suggest that:
H2: The numerical star rating of online reviews has an inverse relationship with the review usefulness.
Source Credibility
Information available in profiles of online community users, especially those on
online consumer review sites, provides valuable information about the message source (Baek
et al. 2012; Forman et al. 2008; Kuan et al. 2015). Source credibility, a recipient’s perception
of the credibility of a reviewer, is an important ELM peripheral cue that has been studied in
the past (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009; Kuan et al. 2015). Source
credibility has been shown to have a positive effect on the credibility of eWOM in online
consumer review platforms (Cheung et al. 2012; Cheung et al. 2009). Past studies have also
found source credibility to have a positive effect on the usefulness of online reviews. Forman
et al. (2008) found that revealing the identity of a reviewer has a positive effect on the
usefulness of reviews. More recent studies (Baek et al. 2012; Kuan et al. 2015) have also
found that platform-based signals of source credibility, such as the top reviewer badge on
Amazon, have a positive effect on the usefulness of online reviews. As Yelp uses a
comparable feature in awarding an Elite badge to its top reviewers, we argue that Yelp’s top
reviewer badge is a good proxy of reviewer credibility. Thus, we posit that:
H4: Reviews from Elite badge members are more likely to receive useful votes than those without badges.
Review Consistency
Past studies have examined the effect consistency of online reviews on review usefulness. (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009; Quaschning et al.
53


2015; Zhang and Watts 2003). These prior studies examine the extent to which a recommendation by a reviewer is consistent with those of other reviewers concerning the same product or service evaluation. Baek et al. (2012) have evaluated the consistency of reviews in terms of the agreement between a reviewer’s rating and the average aggregated rating for the associated product/service and have found that a conflict between an individual rating and the aggregated rating has a negative impact on the usefulness of an online review.
While these prior studies have examined the consistency across reviews and in terms of agreement between two peripheral cues (i.e., a reviews rating and the average rating), less attention has been paid to the role of consistency between a review text and the corresponding numerical rating (i.e., the central and peripheral cues of online reviews) and its effect on review usefulness. According to Tang and Guo (2015), eWOM communication begins when a consumer develops attitudes toward a product/service based on his/her consumption experience. The consumer then incorporates those attitudes into a form of review through both textual commentaries (i.e., a central cue) and an associated numerical star rating (i.e., peripheral cue). This transference of consumers’ attitudes into cues contained in the eWOM text is a communication process called encoding or expression. On the other hand, decoding or impression process happens when consumers read online reviews and make the decision whether to accepting or reject the review by evaluating both verbal (i.e., review text) and non-verbal (i.e., review rating) features of online reviews. Since both review text and review rating encapsulate the reviewer’s subjective perspective about a product or service, reading the review text itself should lead the consumers to the same numerical rating during the decoding process (Mudambi et al. 2014). Therefore, the inconsistency between review text and the corresponding review rating is considered as a source of conflicting
54


information for the reader, resulting in ambivalent feelings for the reader (Chang 2012;
Chang 2013; Chang 2016; Monteith 1996), triggering discomfort (Nordgren et al. 2006) and consequently resulting in a lower perceived credibility of the review. Therefore, we suggest that:
H4: The consistency between review text and its attendant star rating has a positive
effect on review usefulness.
Research Methodology
Data Collection
Data in this study was collected from yelp.com. Yelp has been widely recognized as one of the major online consumer review sites on service industries, especially restaurants. We used Data Toolbar, a commercial web scraping software, to collect the consumer reviews. Following Salehan and Kim (2014), we selected restaurants that have received at least 100 reviews. In total, 2,634 usable reviews were collected from 20 restaurants in the metropolitan area of one of the major cities in the western United States in 2015.
Measures
The dependent variable in this study is review usefulness. It is measured as the number of useful votes a review received. The independent variables of this study are review sentiment polarity, review rating, source credibility and review consistency. In addition to them, control variables include the review length, review longevity, average restaurant rating, restaurant’s number of reviews, reviewer’s number of reviews, reviewer’s number of friends, reviewer’s number of followers as well as rating inconsistency.
Measure of the Review Consistency
Review consistency in this study is measured by using two different methods. As the first method, we employ the unsupervised classification method to examine the consistency
55


between the polarity of the review sentiment and the number of stars attached to the review. As the second method, we employ the machine-learning techniques to predict the consistency of the review. These methods are described in detail below.
Method 1: Unsupervised Classification
Generally, there are two main approaches to the problem of extracting sentiment automatically (Pang and Lee 2008). The first method, the lexicon-based (unsupervised) approach, involves calculating the polarity of sentiment based on the words and phrases used in the text using natural language processing and computational linguistic approaches. This is in contrast to the second method, the supervised approach, wherein researchers train classifier algorithms based on predefined labels of the review sentiment (Taboada et al.
2011). Sentiment mining is usually done using the lexicon-based approach because of its efficiency and scalability (Bai 2011). In this study, we used Semantria, which has also been used in past studies (Charissiadis and Karacapilidis 2015; Kim et al. 2015; Lawrence 2014). Semantria is a commercial software that provides an Excel plug-in for conducting an automatic sentiment analysis on text data.4 Semantria breaks each document into different Part of Speech (PoS) and then identifies the sentiment-bearing phrases, which earn a logarithmic scale ranging between -1 and 1. Semantria combines the scores of those phrases to determine the overall sentiment polarity of the document, which can be labeled as either positive, negative, or neutral. In our study, we treated review consistency as the consistency between a review’s numerical star rating and the polarity of sentiment expressed in the review text. We labeled reviews that received 1 or 2 star(s) as negative, 3 stars as neutral, and 4 or 5 stars as positive. The review consistency variable was treated as a binary variable; that
4 https://semantria.com/support/resources/technology
56


is, it received a “1” if the review sentiment polarity and review star rating labels were equal, and “0” otherwise.
Method 2: Supervised Classification
As the second method to measure the review consistency, a machine-learning technique (a supervised approach) was used, which involves training of classifier algorithms. To perform this supervised classification, the Graphlab module in Python was used. Graphlab is an extensible machine-learning framework that includes rich libraries for data transformation and manipulation as well as a task-oriented machine-learning toolkit for creating, evaluating, and improving the machine-learning models. Before training the classifiers, we parsed the text data to remove unwanted characters and punctuation. To train the classifiers, a “term-by-frequency matrix” was created. In the term-by-frequency matrix, each column represents a unique word that appears across all reviews, and each row refers to each individual review text. Each cell in the matrix represents the number of times that a term (column) appears in a particular review text (row). Since somewhat meaningless words such as “a,” “all,” or “the” in the reviews are not useful in influencing judgment, they were eliminated from the matrix in order to reduce the number of columns.
While term-by-frequency matrix weighting has been used to train classifier algorithms in past studies (Ngo-Ye and Sinha 2014), using term frequency alone cannot effectively distinguish among reviews (Cao et al. 2011) because a term that appears commonly in one type of review may also appear in other types. For instance, in Yelp reviews, the term “food” may appear in many reviews, encompassing reviews that receive many “useful” votes as well as reviews that do not receive any votes. In order to address this problem, the term frequencies were adjusted by TF-IDF weighting (term frequency-inverse
57


document frequency). A term’s TF-IDF value increases comparably to the number of times a word appears in a review, but it is offset by the frequency of the word in all reviews (Arazy and Woo 2007; Cao et al. 2011; Salton et al. 1975). The TF-IDF has been reported to increase information retrieval precision by up to 70% when compared to the term-by-frequency matrix (Salton et al. 1975). Equation 1 expresses the standard TF-IDF weighting (Arazy and Woo 2007; Cao et al. 2011; Salton et al. 1975):
wij = tfijXidfi
In the above equation, wtj is the weighted frequency of term i in document j; tfcj is the frequency of term i in document j; and id/\ = log2(iV/nj) + 1, where JV is the number of the documents and is the frequency of term i in the documents.
In the next step, 400 records were randomly selected to train and test the performance of the classifiers. Two undergraduate students were asked to read the online reviews and examine the consistency between the review text and the associated numerical star rating.
The coders were asked to assign “1” to the reviews if they believed the information provided in the review text is consistent with the numerical star rating of the review, and “0” otherwise. The Cohen's kappa (Carletta 1996) inter-rater reliability measure was found to exceed 0.7, indicating a probability that the agreed understanding between the two student coders was significantly higher than what can be obtained by chance (Krippendorff 2012; Landis and Koch 1977). The coders next met to resolve conflicting ratings until the overall agreements converged. The review consistency labels provided through the coding process were used to train and test the performance of the classifiers in the machine-learning step.
In addition, we sought to evaluate the performance of support vector machines (SVMs) and random forest classifiers in predicting the review consistency labels. These
58


classifiers have previously been used to study the usefulness of online reviews (Ghose and Ipeirotis 2011). To predict the review consistency labels, the 400 selected records were randomly split into 70% of training data and 30% of test data to test their performance. The TF-IDF matrix and the numerical star rating of the reviews were used as the set of features to train the classifiers based on the training set. Further, accuracy metrics were used to evaluate the performance of the classifiers based on the test set. Our results show that the random forest algorithm (with 91% accuracy) outperforms the SVM classifier (with 61% accuracy), which is consistent with the results of the past study (Ghose and Ipeirotis 2011).
Finally, the model built on random forest classifiers was used to predict the unknown review consistency labels in the remaining 2234 records. Those 2234 records were later used to empirically test the hypotheses using ordered logit and negative binomial regression. Figure 3 summarizes the steps involved in predicting the review consistency using the machine-learning approach.
Other Measures_______
Source credibility, which refers to the credibility of the reviewer (Baek et al. 2012; Kuan et al. 2015), is operationalized as a binary variable. If the reviewer was an Elite badge member, the source credibility was assigned as “1”; otherwise, “0”. Review length was measured as the number of words contained in the review (Baek et al. 2012; Salehan and Kim 2014; Salehan and Kim 2016). Review longevity was measured by the number of months elapsed from the date when the review was posted. Rating inconsistency was measured by the difference between a review’s star rating and the average numerical star ratings of the corresponding restaurant. The review sentiment polarity, ranging from -1 to 1, was used to understand whether a review was negative, neutral or positive. In addition,
59


review longevity and review, other variables such as restaurant’s number of reviews, reviewer’s number of reviews, reviewer’s number of tips, reviewer’s number of friends and reviewer’s number of followers are included as control variables in our study.
Empirical Models
In this study, we are interested in understanding (1) why a review receives a useful vote, and (2) what factors contribute for reviews to receive useful votes (Cao et al. 2011).
The ordered logit model, as an extension of the binary logit model, can accommodate a dependent variable that has more than two categories. In our study, the dependent variable is the number of useful votes a review has received from other online users. We followed the Cao et al. (2011) approach to categorizing the number of useful votes. Table 4 shows the reviews by the number of useful votes in the eight categories.
Moreover, the dependent variable of our study is the number of useful votes received by other online users. As is shown in Table 4, a large proportion of the reviews (50.21%) did not receive a single useful vote, which shows the overdispersed variance of useful votes. Therefore, we used negative binomial regression, one of the Poisson model variations (Greene 1994; Schindler and Bickart 2012) . To ensure that negative binomial regression was suitable for the purposes of this study over the Poisson model, we tested whether the overdispersion parameter a was significantly different from zero. Our results gave a p-value < 0.001, which confirms the existence of overdispersion in our dataset, validating that it was appropriate to use negative binomial regression over the Poisson model. Therefore, our model was estimated by using both an ordered logit and a negative binomial regression. We built 5 models, in order to check whether adding a new variable of review consistency significantly improves the variance of the review usefulness. The basic model (model 1 in
60


Table 5) contains only control variables. Model 2 estimates the full model excluding the review consistency variables. Model 3 estimates the full model including the unsupervised variable of the review consistency. Model 4 estimates the full model including the supervised variable of the review consistency, and model 5 estimates the full model including both the unsupervised and supervised variables of the review consistency. The below equations show our regression models. In all the equations, the variables in italics are log-transformed.
Model 1: Control variables:
Usefulness = (30 + /?i \n(Review Length) + (32 \n(Review Longevity) +
/?3(Average Restaurant Rating) + (34\n(Restaurant's Number of Reviews) +
(35 (Rating Inconsistency) + /?6ln (Reviewer’s Number of Reviews) +
(37 \n(Reviewer's Number of Friends) + (38 \n(Reviewer's Number of Tips) + (39\n(Reviewer's Number of Followers) + e
Model 2: All the variables excluding review consistency variables:
Usefulness = (30 + f>i \n(Review Length) + (32 \n(Review Longevity) +
/?3(Average Restaurant Rating) + (]4\n(Restaurant's Number of Reviews) +
(3$ (Rating Inconsistency) + (3e In (Reviewer’s Number of Reviews) +
(37 ln(/?etnewer's Number of Friends) + (38 \n(Reviewer's Number of Tips) +
(39 \n(Reviewer's Number of Followers) + /?10(Review Sentiment Polarity) +
/?n (Review Rating) + (312(Elite Badge Memeber) + e
Model 3: All the variables including the unsupervised measure of the review consistency variable:
Usefulness = (30 + (3X \n(Review Length) + (32 \n(Review Longevity) +
(33(Average Restaurant Rating) + (34\n(Restaurant's Number of Reviews) +
61


/?5 (Rating Inconsistency) + /?6ln (Reviewer’s Number of Reviews) +
/37 \n(Reviewer's Number of Friends) + (38 \n(Reviewer's Number of Tips) +
/?9 ln(Rei;iewer's Number of Followers) + /?10(Review Sentiment Polarity) +
/?n (Review Rating) + /?12(Elite Badge Memeber) +
/?13 (Unsupervised Review Consistency) + e
Model 4: All the variables including the supervised measure of the review consistency variable:
Usefulness = (30 + /?i \n(Review Length) + f>2 \n(Review Longevity) +
/?3(Average Restaurant Rating) + (34\n(Restaurant's Number of Reviews) +
(3$ (Rating Inconsistency) + /?6ln (Reviewer’s Number of Reviews) +
/?7 ln(Rei;tewer's Number of Friends) + /?8 ln(Rei;tewer's Number of Tips) +
/?9 ln(Reinewer's Number of Followers) + /?10(Review Sentiment Polarity) +
/?n (Review Rating) + /?12(Elite Badge Memeber) +
/?13 (Supervised Review Consistency) + £
Model 5: All the variables including the unsupervised and supervised measures of the review consistency variable:
Usefulness = /?0 + f>i \n(Review Length) + /?2 \n(Review Longevity) +
/?3(Average Restaurant Rating) + (]4\n(Restaurant’s Number of Reviews) +
(3$ (Rating Inconsistency) + /?6ln (Reviewer’s Number of Reviews) +
/?7 ln(Rei;tewer's Number of Friends) + (38 \n(Reviewer's Number of Tips) +
/?9 ln(Reinewer's Number of Followers) + /?10(Review Sentiment Polarity) +
/?n (Review Rating) + /?12(Elite Badge Memeber) +
(313(Unsupervised Review Consistency) + /?14(Supervised Review Consistency) + e
62


Discussion
Models 3 column in Table 4 is used to evaluate our hypotheses. The first hypothesis (Hi) examined the effect of review sentiment polarity on the review usefulness. We expected that a negative sentiment polarity would positively influence the review usefulness.
However, we did not find support for this hypothesis. One possible explanation for this finding might be related to the effect of argument quality of review usefulness (Cheung et al. 2012; Cheung et al. 2009). In other words, it is possible that a review with low numerical ratings (i.e., one or two-star ratings) gains the attention of consumers and persuades them to read the review. However, in the review text, the reviewer doesn’t provide convincing arguments about why he/she assigns a low numerical rating to the restaurant. In other words, the reviewer might be anxious or angry and tends to express his/her emotions by writing a review. However, his/her emotions weaken the rational arguments in the review text (Wang 2006), resulting in making his/her review less useful.
Our second hypothesis (H2) examined the effect of the review rating on the review’s usefulness. Our results show that the numerical star rating of a review has a negative effect on the usefulness of online reviews (/?n= -0.17, p<0.01), and its exponentiated odds ratio5 is
5 To better interpret the results of ordered logit in our analysis we used the value of odds ratio. The odds ratio is better than /? coefficient because it doesn’t require logarithmic transformation. The odds ratios can be obtained by exponentiating the ordered logit coefficients (i.e., exp(/?)) (Field 2009). The odds ratio is defined as: Aodds=--------------------------------------. Given all other predictors held constant, a value greater than 1
original odds
indicates that as the predictor increases by one unit, the odds of outcome increases. On the other hand, a value less than one indicates that as the predictor increases by one unit, the odds of outcome occurring decreases. In other words, the negative value of /? indicates a decrease in the likelihood of the outcome occurring given one unit increase in the predictor and a positive value of /? indicates an increase in the likelihood of the outcome occurring given one unit increase in the predictor. F or instance, in HI. /?,,= -0.17 and the odds ratio is represented as xp(-0.17) ) ~ 0.84, indicating 16% decrease in the likelihood of receiving a useful vote. On the other hand, the values of /? coefficient is positive in H2, H3 and H4, indicating an increase in the likelihood of receiving useful votes given one unit increase in the predictors.
63


represented as exp(—0.17) « 0.84. That means, given the other variables in the model held constant, for a one unit increase in review numerical rating, the likelihood of receiving useful votes decreases by 16%. Our finding confirms the existence of a negativity bias among online reviews (Kuan et al. 2015). As reported by (Gao et al. 2015b; Hu et al. 2009), the majority of online reviews are positive, which reflects the typical J-shaped distribution of star rating. Consistently, in our dataset, most of the reviews (70%) were found to have either four or five star ratings (refer to Figure 4). Thus, since most of the reviews in our dataset are positive, negativity bias suggests that negative information is generally deemed more diagnostic than positive information (Kuan et al. 2015).
The third hypothesis (H3) contemplated the effect of source credibility on review usefulness. Our results show that a review by an Elite badge member is more likely to receive a usefulness vote (/?i2=0.84, p<0.01). The odds ratio is exp(0.84) « 2.31.That means, given all other variables held constant, an elite-badge member is 2.31 times more likely to receive a useful vote than a non-elite badge member. Thus, H3 was supported. Our results also confirm the findings of past studies (Baek et al. 2012; Kuan et al. 2015) reporting the positive effect of platform-generated measures of source credibility (i.e., top reviewer rank on Amazon) on review usefulness.
Our fourth hypothesis (H4) examined the effect of review consistency on the review usefulness. In model 3, we included the unsupervised measure of review consistency and examined its effect on the review usefulness. The result of the ordered logit regression indicates a positive effect of the unsupervised measure of review consistency on review usefulness (/?13=0.78, p<0.01). The odds ratio is exp(0.78) « 2.18. That means, given all other variables held constant, consistent reviews are 2.18 times more likely to receive useful
64


votes than inconsistent reviews. In model 4, we examined the effect of supervised measure of review consistency on review usefulness, and our results support its positive effect on review usefulness (/?13=0.83, p<0.01). Therefore, our models show strong support for H4.
In model 5, we included both the unsupervised and supervised variables of review consistency. Our result yields the existence of multicollinearity problem, because of the high correlation between the unsupervised and supervised variables of review consistency. To address the problem, we further examine the variance inflation factor (VIF) of the independent variables based on model 5, and the result shows high VIF values for Unsupervised Review Consistency (VIF=5.2) and Supervised Review Consistency (VIF=8.3). To check whether adding the review consistency variable in model 3 would enhance the R2 above and beyond model 1 and model 2, we run a hierarchical regression on the successive models. As presented in Table 5, our result shows that adding the review consistency variable to model 3 significantly improves the variance in the usefulness of online reviews above and beyond model 1 and model 2. However, this result does not hold for model 4 and model 5 because of the high correlation between supervised and unsupervised measures of review consistency. Given the focal variable (i.e., review consistency) that we developed and tested in this study, it confirms the validity of ELM in the online review context. In other words, online reviewers tend to process central cue (i.e., verbal review text) and peripheral cue (i.e., non-verbal star rating) simultaneously, not in isolation.
Post-hoc Analysis
Our post-hoc analysis is theoretically motivated by two important reasons. First, prior studies on eWOM communication process have maintained that the credibility of a reviewer
65


(e.g., if the reviewer is an elite badge member or no) is the most important peripheral cue that affects eWOM credibility. Likewise, Baek et al. (2012) and Kuan et al, (2015) find that the platform generated signal of reviewers’ credibility such as top reviewer in Amazon has a positive effect on review usefulness. Second, ELM states that individuals start to process information by incorporating peripheral cues, and then, depending on the quality or strength of peripheral cues, they might be motivated to process central cues (Mousavizadeh et al.
2015; Petty and Cacioppo 2012; Sundar and Kim 2005).
Reflecting the prior studies and by considering that elite badge member turns out to be a strong peripheral cue as a proxy of the reviewer credibility (refer to Table 5), we believe examining the different structural pattern of review usefulness based on whether a reviewer is credible (i.e., elite badge =1) or not credible (i.e., elite badge =0) will shed more light on how consumers process central and peripheral cues.
As reported in table 6, our analysis reveals that elite badge members and non-elite badge members show different patterns in processing central and peripheral cues. To evaluate online elite badge members’ reviews, online users mainly focus on reading review text (i.e., review length where /?=0.54, exp(0.54) =1.71), and disregard other peripheral cues, because the presence of elite badge itself works as a strong peripheral cue that warrants the credibility of the reviewers. In other words, the existence of elite badge member cue overrides the signals of other peripheral cues, and direct online users’ attention directly to the content of online reviews. Therefore, reviews that elaborate more on a service (i.e., lengthier reviews) are more likely to receive useful votes with an expectation of online users that reviews written by elite badge members would provide credible information nuggets about the service quality of reviewed restaurants (Baek et al. 2012). In contrast, if a review is
66


written by a non-elite badge member, online users would have struggled to process all other peripheral cues to assess if an online review would be credible. It is reflected by statistically significant effects of various peripheral cues (i.e., review rating, rating inconsistency, review longevity, reviewers’ number of friends), including the central cue of the review content (i.e., review length), on the review usefulness.
Interestingly, Table 6 shows that the review consistency variable has different effects on review usefulness. That means, in the case of a review written by an elite badge member, the effect of review consistency turns out to be insignificant. However, if a review is written by a non-elite badge member, the review consistency becomes a significant driver of review usefulness (/?=0.62, exp(0.62) = 1.85). For the different effects of review consistency, we submit two explanations. First, in case of an elite badge member’s review, the insignificant effect of review consistency can be attributable to the sweeping effect of elite badge member as a strong peripheral cue. That means, as the existence of an elite badge itself guarantees the credibility of the review, online users begin to read the reviews even without bothering themselves to evaluate if the review shows consistency between review text and review rating. However, as a review posted by non-elite badge member does not guarantee the credibility of the reviewer, online users would have exerted extra cognitive effort to evaluate if the review shows consistency between review text and review rating. Second, the different effects of review consistency between elite and non-elite member can be related with the different length of review (i.e., average number of words per review). That means, as elite badge members tend to write longer reviews that non-elite badge members6, the former
6 While, average length of reviews (i.e., number of word per review) by elite badge member is 162, the same by non-elite badge members is 89.
67


reviews would have made it harder for online users to evaluate the consistency between review text and review rating.
Theoretical and Practical Contributions Theoretical Contributions
This study contributes to prior review usefulness literature in two ways. First, ELM states that central and peripheral cues of a message are processed jointly, not in isolation, by message recipients. Therefore, in the online review context, it makes more sense to consider review rating (i.e., the peripheral cue) and the review text (i.e., the central cue) simultaneously. However, prior studies (Baek et al. 2012; Forman et al. 2008; Kuan et al. 2015; Mudambi et al. 2014; Salehan and Kim 2014; Yin et al. 2014a) have focused on single factor measures (i.e., either review rating or review text) to explain their isolated effect on review usefulness. This study fills this theoretical gap by crafting and operationalizing a two-factor variable (i.e., review consistency), which measures the consistency between review text and review rating (i.e., central and peripheral cues) and its effect on review usefulness. Our analysis empirically demonstrates that our review consistency variable better explains the mechanism of online review usefulness above and beyond prior studies.
Second, while prior research suggested that reviewer credibility is the most important peripheral cue of ELM, its mechanism has not been well explained (Baek et al. 2012; Cheung et al. 2012; Cheung et al. 2009). However, consistent with ELM, our post-hoc analysis empirically demonstrates the mechanism that peripheral cue (i.e., elite badge) serves as signals that motivate online users to process central cue (i.e., review text) in certain ways (Kuan et al. 2015; Mousavizadeh et al. 2015; Petty et al. 2002). Our results suggest that, in presence of reviewer credibility signal (i.e., an elite-bade member in this study), online users
68


move their focus to review text itself and other peripheral cues become less important. However, in the absence of the credible signal of the elite badge, online users exert extra cognitive efforts to mobilize other central and peripheral cues as a means to evaluate the information quality of online reviews.
Practical Contributions
Online retailers can benefit from the result of this study by incorporating review consistency variable in sorting online reviews. Specifically, the unsupervised measure of review consistency can be integrated into the sorting algorithms of online review systems, because of its low-cost implementation. An alternative approach is to examine the consistency between review text and review rating before a reviewer post a review. That means, the review system can be complemented with a real-time alert component that can inform online reviewers with the possible inconsistency between review text and review rating before he/she concludes the review.
Limitations and Future Research
The sample used in this study is limited to restaurants’ reviews on Yelp. Therefore, findings in this study should be generalized with caution. Previous studies have shown the different effects of platforms (Lee and Youn 2009) and product types (Baek et al. 2012; Mudambi and Schuff 2010) on eWOM. Thus, more research needs to be done to know how different product types and platforms would affect the review usefulness vote.
This study also opens a new avenue for how the text mining techniques can be used to examine the usefulness of online reviews. For instance, the argument has been named as one the most important factors that influence the perceived credibility of online reviews (Cheung et al. 2012; Cheung et al. 2009). Prior studies have measured argument quality as a latent
69


variable by using data collected through surveys or experiments (Awad and Ragowsky 2008; Cheung et al. 2012; Cheung et al. 2009; Chu and Kamal 2008; Park et al. 2007). These prior measure argument quality as a construct that reflects the degree to which an online review is believable, strong, persuasive, good and informative (Cheung et al. 2012; Cheung et al. 2009). One limitation of this view of argument quality is it is difficult in practice for online review providers to quantify these dimensions so that they can improve their recommendation systems. We believe, an alternative approach to quantifying argument quality is to draw on the studies that suggest more quantifiable dimensions of argument quality. For instance, the recommendation consistency (Baek et al. 2012) can be examined as the degree of consistency between the textual content of an online review and other reviews concerning the same product or service. In addition, the review accuracy (Otterbacher 2009) can be studied as the degree of consistency between a review and the description of the corresponding product on online consumer review sites. Therefore, it would be fruitful for future studies to use text mining and natural language processing to examine the effect of these variables on the review usefulness.
Concluding Remarks
Quality online reviews play an important role in helping online consumers make an informed decision about their purchase. Thus, understanding the determinant factors of useful online reviews is important for both scholars and online consumer review providers. By using ELM as the theoretical lens and using machine learning and sentiment mining in combination with econometrics analysis, we crafted and operationalized a two-dimensional variable (i.e., review consistency), which measures the consistency between review text and review rating, and its effect on review usefulness. In theory, this study advances our
70


understanding of ELM in online review context by providing empirical evidence that the two-factor based review consistency variable outperforms the single factor variables of previous studies in explaining how online reviews are considered as useful by online users.
In addition, our results shed light on how consumers process central and peripheral cues in different ways to evaluate online reviews based on whether or not an online review is written by an elite badge member.
71


FIGURES AND TABLES
Derek M.
Seattle, WA
â–¼ â–¼4 friends Q 60 reviews
Elite ’16
0 O O Q Q 7/8/2016
1 love variety, and fortunately at Umma's variety is the name of the game. For my first visit I got the larger box in order to try a little bit of everything, to then narrow down my favorites for next time. The problem.. .was that the majority of them proved to be favorite-worthy. It was a great combination of Korean favorites I've come to love, with new things like spam musubi that I hadn't had a chance to try until now.
The other great quality of the food was the freshness. All veggies were green and crisp; all meat I had was tender and warm.
Will totally be back soon, and will be sure to bring my box back to get $1 off.
Was this review ...?
© Funny ("■) Cool 1
(J) Useful 1
Figure III-1. An example of a Review on Yelp
Figure III-2. Research Model
72


Number of reviews
f Step 1 Text Parsing \
\ /
Step 2
Term Reduction (stop words list)
Step 3
Random Split the Main Dataset N=2634
Step 4
Machine Learning Steps
Removing Unwanted Characters and
Punctuations
Words Per Sentence
Removing Stop Words (i.e., words that do not have any meaning such as "a", "the" etc.) Term-Frequency-Inverse-Document-Frequency (TF-IDF)
Split 1:
To Test the Research Model N=2234
Split 2:
To Train and Test the Performance of the Classifiers N=400
Divide Split 2 randomly into the training set 970%) and the test set (30%)
Train Random Forest and Support Vector Machine based on the training set Evaluate the accuracy of the methods on the test set
Use the best Model (Random Forest) to Predict the Review Consistency Labels in the Split 1
Figure III-3. Supervised Measure of the Review Consistency
900
800
700
600
500
400
300
200
100
0
845

724



323
228
114
1 2 3 4 5
Numerical Star Ratings
Figure HI-4. The Distribution of Rating Based on Number of Reviews
73


Table III-1. Descriptive Statistics
Variables Description N Mean St. Dev Median Min Max
USEFULNESS Number of useful votes a review 2,234 0.98 1.6 0 0 20
Elite Badge member Whether a reviewer is an Elite 2,234 0 0.27 0 0 1
Unsupervised Review Unsupervised measure of the recommendation consistency 2,234 0.59 0.49 1 0 1
Supervised Review Consistency Supervised measure of the recommendation consistency 2,234 0.52 0.5 1 0 1
Restaurant’s Number of Reviews Number of reviews a restaurant has received 2,234 523.63 245.20 531 155 942
Average Restaurant Rating Average numerical star rating of a restaurant 2,234 3.83 0.38 4 3 4
Reviewer’s Number of Friends Reviewer’s number of friends on Yelp 2,234 63.12 119.60 15 0 1,12
Reviewer’s Number of Followers Reviewer’s number of followers on Yelp 2,234 2.54 7.61 0 0 84
Reviewer’s Number of Reviews Number of reviews posted by a reviewer on Yelp 2,234 131.96 238.70 32.5 1 2,83
Reviewer’s Number of Tips Number of tips posted by a reviewer on Yelp 2,234 22.33 76.23 1 0 1,32
Review Longevity Months elapsed after a review has been posted 2,234 25.33 20.27 19 0 111
Review Length Length of review text 2,234 108.69 98.17 80 2 953
Review Rating Numerical star ratings of a review 2,234 3.86 1.17 4.00 1 5
Rating Inconsistency The absolute difference between numerical star rating of review and the average star rating of the corresponding restaurant 2,234 0.84 0.71 1 0 3
Review Sentiment Polarity Review sentiment polarity 2,234 0.49 0.68 1 -1 1
74


Table III-2. Correlation Matrix
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1-USEFULNESS 1.00
2- Elite Badge Member 0.39 1.00
3- Unsupervised Review Consistency 0.09 0.01 1.00
4- Review Length 0.41 0.33 -0.02 1.00
5- Average Restaurant Rating 0.04 0.14 0.15 0.17 1.00
6- Restaurant’s Number of Reviews -0.03 0.13 0.01 0.12 0.42 1.00
7- Review Longevity 0.22 0.05 -0.06 0.08 -0.36 -0.26 1.00
8- Reviewer’s Number of Reviews 0.38 0.56 0.00 0.28 0.08 0.07 0.18 1.00
9- Reviewer’s Number of Friends 0.42 0.52 0.03 0.30 0.09 0.09 0.16 0.65 1.00
10- Reviewer’s Number of Tips 0.27 0.29 0.03 0.15 0.02 0.03 0.13 0.44 0.44 1.00
11- Reviewer’s Number of Followers 0.26 0.37 -0.02 0.21 0.01 0.06 0.13 0.59 0.51 0.22 1.00
12- Supervised Review Consistency 0.06 0.04 0.87 -0.03 0.23 0.08 -0.11 0.00 0.05 0.04 -0.01 1.00
13 - Review Rating -0.05 0.07 0.33 -0.07 0.32 0.09 -0.10 0.03 0.06 0.03 0.01 0.51 1.00
14- Rating Inconsistency 0.03 -0.16 -0.09 0.01 -0.12 -0.02 -0.05 -0.14 0.13 -0.07 -0.09 -0.18 -.0.38 1.00
15- Review Sentiment Polarity 0.03 0.07 0.44 -0.00 0.15 0.03 -0.05 0.04 0.07 0.05 0.02 0.64 0.41 -0.25 1.00
Table III-3. Categories of Useful Votes
Number of useful votes Number of reviews Percentage (%)
0 1124 50.21
1 916 27.70
2 244 10.92
3 116 5.19
4 61 2.73
5 25 1.11
6 19 0.85
7 or more 26 1.16
75


Table III-4. Econometric Analyses, Ordered Logit
Variables Model 1 Model 2 Model 3 Model 4 Model 5
Elite Badge Member 0.85*** 0 84*** 0.82*** 0.83***
Review Rating -0.10** -0 17*** -0.22*** -0 19***
Review Sentiment Polarity 0.29** 0.1 -0.04 0.03
Un supervised Review Consistency 0 yg*** 0.53***
Supervised Review Consistency 0.83*** 0.3
Review Length 0.53*** 0.50*** 0.50*** 0.51*** 0.51***
Average Restaurant Rating 0.1 0.1 0.07 0.05 0.06
Restaurant’s Number of Reviews -0 98*** _q 99*** -0 94*** -1.02*** -0 97***
Review Longevity 022*** 0.26*** 0 27*** q 27*** q 27***
Rating Inconsistency 0.39** q 4Q*** 0 27*** q 37*** q 37***
Reviewer’s Number of Reviews 0.12** 0.03 0.05 0.05 0.05
Reviewer’s Number of Friends 0.15** 0.13*** 0.12*** 0.12*** 0.12***
Reviewer’s Number of Tips 0.05** 0.04* 0.04* 0.04* 0.04*
Reviewer’s Number of Followers 0.04 0.01 0.02 0.02 0.02
Observation 2,234 2,234 2,234 2,234 2,234
R2 0.3 0.32 0.35 0.34 0.35
Chi2 744.85*** 805.10*** 865.87*** 860.04*** 868.28***
Note: *p<0.1; **p<0.05; ***p<0.0
76


Table III-5. Comparing Successive Models
RSS DF Sum of Square F-statistics P-value
Model 1: Control Variables 3015.7
Model 2: Model 1 +Independent Variables (Excluding Review Consistency Variables) 2957.5 3 58.19 14.81 <0.001
Model3: Model 2+Unsupervised Review Consistency 2910.8 1 46.69 35.65 <0.001
Model 4: Model 2+Supervised Review Consistency 2908.3 0 2.538
Model 5: Model 2+Unsupervised Review Consistency+ Supervised Review Consistency 2905.9 1 2.418 1.8464 0.1743
Note: *p<0.05; **p<0.01; ***p<0.001
Table III-6. Post-hoc Analysis
Variables Elite Badge Members Non-Elite Badge Members
Loait Loait
Review Rating -0.22 -0.14**
Review Sentiment Polarity -0.32 0.04
Unsupervised Review Consistency 20.65 0.62***
Review Length 0.54*** 0.38***
Average Restaurant Rating 0.13 0.08
Restaurant’s Number of Reviews -1.93** -0.68***
Review Longevity 0.13 0.31***
Rating Inconsistency 0.31 q 29***
Reviewer’s Number of Reviews -0.13 0.004
Reviewer’s Number of Friends 0.17 0.10***
Reviewer’s Number of Tips 0.06 0.01
Reviewer’s Number of Followers 0.03 0.13*
Observation 596 1,638
Note: *p<0.1; **p<0.05; ***p<0.01
77


REFERENCES
Adams, D. A., Nelson, R. R., and Todd, P. A. 1992. "Perceived Usefulness, Ease of Use, and Usage of Information Technology: A Replication," MIS quarterly (16:2), pp. 227-247.
Angst, C. M., and Agarwal, R. 2009. "Adoption of Electronic Health Records in the Presence of Privacy Concerns: The Elaboration Likelihood Model and Individual Persuasion," MIS quarterly (33:2), pp. 339-370.
Arazy, O., and Woo, C. 2007. "Enhancing Information Retrieval through Statistical Natural Language Processign: A Study of Collocation Indexing," MIS quarterly (24:2), pp. 525-546.
Arnold, H. J., and Feldman, D. C. 1981. "Social Desirability Response Bias in Self-Report Choice Situations," Academy of Management Journal (24:2), pp. 377-385.
Awad, N. F., and Ragowsky, A. 2008. "Establishing Trust in Electronic Commerce through Online Word of Mouth: An Examination across Genders," Journal of Management Information Systems (24:4), pp. 101-121.
Baek, H., Ahn, J., and Choi, Y. 2012. "Helpfulness of Online Consumer Reviews: Readers' Objectives and Review Cues," International Journal of Electronic Commerce (17:2), pp. 99-126.
Bai, X. 2011. "Predicting Consumer Sentiments from Online Text," Decision Support Systems (50:4), pp. 732-742.
Bakshy, E., Eckles, D., Yan, R., and Rosenn, I. 2012. "Social Influence in Social Advertising: Evidence from Field Experiments," Proceedings of the 13th ACM Conference on Electronic Commerce: ACM, pp. 146-161.
Barclay, D., Higgins, C., and Thompson, R. 1995. "The Partial Least Squares (Pis) Approach to Causal Modeling: Personal Computer Adoption and Use as an Illustration," Technology studies (2:2), pp. 285-309.
Baumeister, R. F., Bratslavsky, E., Finkenauer, C., and Vohs, K. D. 2001. "Bad Is Stronger Than Good," Review of general psychology (5:4), p. 323.
Berger, J., and Milkman, K. L. 2012. "What Makes Online Content Viral?," Journal of marketing research (49:2), pp. 192-205.
Bhattacheijee, A., and Sanford, C. 2006. "Influence Processes for Information Technology Acceptance: An Elaboration Likelihood Model," MIS quarterly (30:4), pp. 805-825.
Brown, J. J., and Reingen, P. H. 1987. "Social Ties and Word-of-Mouth Referral Behavior," Journal of Consumer research), pp. 350-362.
78


Cao, Q., Duan, W., and Gan, Q. 2011. "Exploring Determinants of Voting for the “Helpfulness” of Online User Reviews: A Text Mining Approach," Decision Support Systems (50:2), pp. 511-521.
Carletta, J. 1996. "Assessing Agreement on Classification Tasks: The Kappa Statistic," Computational linguistics (22:2), pp. 249-254.
Carlson, K., and Guha, A. 2010. "The Ratings Paradox: Why We Prefer Reading Negative Reviews, but Then Subsequently Rate These Reviews as Less Useful," Advances in Consumer Research (37).
Casteleyn, J., Mottart, A., and Rutten, K. 2009. "How to Use Data from Facebook in Your Market Research," International Journal of Market Research (51:4), pp. 439-447.
Chang, C. 2012. "Ambivalent Attitudes in a Communication Process: An Integrated Model," Human Communication Research (38:3), pp. 332-359.
Chang, C. 2013. "The Influence of Ambivalence toward a Communication Source: Media Context Priming and Persuasion Polarization," Communication Research), p. 0093650213495408.
Chang, C. 2016. "Responses to Conflicting Information in Computer-Mediated Communication: Gender Difference as an Example," new media & society (18:1), pp. 5-24.
Charissiadis, A., and Karacapilidis, N. 2015. "Strengthening the Rationale of Recommendations through a Hybrid Explanations Building Framework," in Intelligent Decision Technologies. Springer, pp. 311-323.
Cheung, C. M.-Y., Sia, C.-L., and Kuan, K. K. 2012. "Is This Review Believable? A Study of Factors Affecting the Credibility of Online Consumer Reviews from an Elm Perspective," Journal of the Association for Information Systems (13:8), p. 618.
Cheung, C. M., Lee, M. K., and Rabjohn, N. 2008. "The Impact of Electronic Word-of-Mouth: The Adoption of Online Opinions in Online Customer Communities," Internet Research (18:3), pp. 229-247.
Cheung, M. Y., Luo, C., Sia, C. L., and Chen, H. 2009. "Credibility of Electronic Word-of-Mouth: Informational and Normative Determinants of on-Line Consumer Recommendations," International Journal of Electronic Commerce (13:4), pp. 9-38.
Chevalier, J. A., and Mayzlin, D. 2006. "The Effect of Word of Mouth on Sales: Online Book Reviews," Journal of marketing research (43:3), pp. 345-354.
Chin, W. W. 1998a. "Commentary: Issues and Opinion on Structural Equation Modeling." JSTOR.
79


Chin, W. W. 1998b. "The Partial Least Squares Approach to Structural Equation Modeling," Modern methods for business research (295:2), pp. 295-336.
Chin, W. W., Marcolin, B. L., and Newsted, P. R. 2003. "A Partial Least Squares Latent Variable Modeling Approach for Measuring Interaction Effects: Results from a Monte Carlo Simulation Study and an Electronic-Mail Emotion/Adoption Study," Information systems research (14:2), pp. 189-217.
Chu, S.-C., and Choi, S. M. 2011. "Electronic Word-of-Mouth in Social Networking Sites: A Cross-Cultural Study of the United States and China," Journal of Global Marketing (24:3), pp. 263-281.
Chu, S.-C., and Kamal, S. 2008. "The Effect of Perceived Blogger Credibility and Argument Quality on Message Elaboration and Brand Attitudes: An Exploratory Study," Journal of Interactive Advertising (8:2), pp. 26-37.
Chu, S.-C., and Kim, Y. 2011. "Determinants of Consumer Engagement in Electronic Word-of-Mouth (Ewom) in Social Networking Sites," International journal of Advertising (30:1), pp. 47-75.
Connors, L., Mudambi, S. M., and Schuff, D. 2011. "Is It the Review or the Reviewer? A Multi-Method Approach to Determine the Antecedents of Online Review Helpfulness," System Sciences (HICSS), 2011 44th Hawaii International Conference on: IEEE, pp. 1-10.
Cooper, S. 2015. "Facebook Is Leading the Way in Social Commerce."
Corson, Y., and Verrier, N. 2007. "Emotions and False Memories Valence or Arousal?," Psychological Science (18:3), pp. 208-211.
Davis, F. D. 1989. "Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology," MIS quarterly), pp. 319-340.
Ebermann, J., Stanoevska-Slabeva, K., and Wozniak, T. 2011. "Influential Factors of Recommendation Behaviour in Social Network Sites-an Empirical Analysis," ECIS.
Edell, J. A., and Burke, M. C. 1984. "The Moderating Effect of Attitude toward an Ad on Ad Effectiveness under Different Processing Conditions," NA-Advances in Consumer Research Volume 11).
eMarketer. 2010. "The Role of Customer Product Reviews,").
eMarketer.com. 2014. "Facebook Is No. 1 for Social Commerce."
Eroglu, S. A., Machleit, K. A., and Davis, L. M. 2003. "Empirical Testing of a Model of Online Store Atmospherics and Shopper Responses," Psychology & Marketing (20:2), pp. 139-150.
80


Fang, Y.-H. 2014. "Beyond the Credibility of Electronic Word of Mouth: Exploring Ewom Adoption on Social Networking Sites from Affective and Curiosity Perspectives," International Journal of Electronic Commerce (18:3), pp. 67-102.
Field, A. 2009. Discovering Statistics Using Spss. Sage publications.
Folkes, V. S. 1988. "The Availability Heuristic and Perceived Risk," Journal of Consumer research (15:1), pp. 13-23.
Forgas, J. P. 1995. "Mood and Judgment: The Affect Infusion Model (Aim)," Psychological bulletin (117:1), p. 39.
Forgas, J. P., and George, J. M. 2001. "Affective Influences on Judgments and Behavior in Organizations: An Information Processing Perspective," Organizational behavior and human decision processes (86:1), pp. 3-34.
Forman, C., Ghose, A., and Wiesenfeld, B. 2008. "Examining the Relationship between Reviews and Sales: The Role of Reviewer Identity Disclosure in Electronic Markets," Information Systems Research (19:3), pp. 291-313.
Fomell, C., and Larcker, D. F. 1981. "Evaluating Structural Equation Models with Unobservable Variables and Measurement Error," Journal of marketing research), pp. 39-50.
Gao, G. G., Greenwood, B. N., Agarwal, R., and Jeffrey, S. 2015a. "Vocal Minority and Silent Majority: How Do Online Ratings Reflect Population Perceptions of Quality?," MIS quarterly (39:3), pp. 565-589.
Gao, G. G., Greenwood, B. N., Agarwal, R., and McCullough, J. S. 2015b. "Vocal Minority and Silent Majority: How Do Online Ratings Reflect Population Perceptions of Quality?,").
Gefen, D., Benbasat, I., and Pavlou, P. 2008. "A Research Agenda for Trust in Online Environments," Journal of Management Information Systems (24:4), pp. 275-286.
Ghose, A., and Ipeirotis, P. G. 2011. "Estimating the Helpfulness and Economic Impact of Product Reviews: Mining Text and Reviewer Characteristics," Knowledge and Data Engineering, IEEE Transactions on (23:10), pp. 1498-1512.
Granovetter, M. S. 1973. "The Strength of Weak Ties," American journal of sociology), pp. 1360-1380.
Greene, W. H. 1994. "Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial Regression Models, ".Leonard N. Stern School of Business, New York Universit).
81


Gupta, P., and Harris, J. 2010. "How E-Wom Recommendations Influence Product Consideration and Quality of Choice: A Motivation to Process Information Perspective," Journal of Business Research (63:9), pp. 1041-1049.
Hansen, S. S., and Lee, J. K. 2013. "What Drives Consumers to Pass Along Marketer-Generated Ewom in Social Network Games? Social and Game Factors in Play," Journal of theoretical and applied electronic commerce research (8:1), pp. 53-68.
Hennig-Thurau, T., Gwinner, K. P., Walsh, G., and Gremler, D. D. 2004. "Electronic Word-of-Mouth Via Consumer-Opinion Platforms: What Motivates Consumers to Articulate Themselves on the Internet?," Journal of interactive marketing (18:1), pp. 38-52.
Herr, P. M., Kardes, F. R., and Kim, J. 1991. "Effects of Word-of-Mouth and Product-Attribute Information on Persuasion: An Accessibility-Diagnosticity Perspective," Journal of consumer research (17:4), pp. 454-462.
Hu, N., Zhang, J., and Pavlou, P. A. 2009. "Overcoming the J-Shaped Distribution of Product Reviews," Communications of the ACM (52:10), pp. 144-147.
Huang, A. H., Chen, K., Yen, D. C., and Tran, T. P. 2015. "A Study of Factors That Contribute to Online Review Helpfulness," Computers in Human Behavior (48), pp. 17-27.
Huang, J., Su, S., Zhou, L., and Liu, X. 2013. "Attitude toward the Viral Ad: Expanding Traditional Advertising Models to Interactive Advertising," Journal of Interactive Marketing (27:1), pp. 36-46.
Jansen, B. J., Zhang, M., Sobel, K., and Chowdury, A. 2009. "Twitter Power: Tweets as Electronic Word of Mouth," Journal of the American society for information science and technology (60:11), pp. 2169-2188.
Kaplan, A. M., and Haenlein, M. 2010. "Users of the World, Unite! The Challenges and Opportunities of Social Media," Business horizons (53:1), pp. 59-68.
Keenan, A., and Shiri, A. 2009. "Sociability and Social Interaction on Social Networking Websites," Library Review (58:6), pp. 438-450.
Kim, J., Naylor, G., Sivadas, E., and Sugumaran, V. 2015. "The Unrealized Value of Incentivized Ewom Recommendations," Marketing Letters), pp. 1-11.
Krippendorff, K. 2012. Content Analysis: An Introduction to Its Methodology. Sage.
Kuan, K. K., Hui, K.-L., Prasarnphanich, P., and Lai, H.-Y. 2015. "What Makes a Review Voted? An Empirical Investigation of Review Voting in Online Review Systems," Journal of the Association for Information Systems (16:1), p. 48.
Landis, J. R., and Koch, G. G. 1977. "The Measurement of Observer Agreement for Categorical Data," biometrics), pp. 159-174.
82


Lawrence, L. 2014. "Reliability of Sentiment Mining Tools: A Comparison of Semantria and Social Mention,").
Lee, J., and Lee, J.-N. 2009. "Understanding the Product Information Inference Process in Electronic Word-of-Mouth: An Objectivity-Subjectivity Dichotomy Perspective," Information & Management (46:5), pp. 302-311.
Lee, J., Park, D.-H., and Han, I. 2008. "The Effect of Negative Online Consumer Reviews on Product Attitude: An Information Processing View," Electronic commerce research and applications (7:3), pp. 341-352.
Lee, M., and Youn, S. 2009. "Electronic Word of Mouth (Ewom) How Ewom Platforms Influence Consumer Product Judgement," International Journal of Advertising (28:3), pp. 473-499.
Lee, W., Xiong, L., and Hu, C. 2012. "The Effect of Facebook Users’ Arousal and Valence on Intention to Go to the Festival: Applying an Extension of the Technology Acceptance Model," International Journal of Hospitality Management (31:3), pp. 819-827.
Lenhart, A. 2009. "Adults and Social Network Websites,").
LePage, E. 2015. "A Long List of Social Media Advertising Stats," in: hootsuite.
Leung, X. Y., Bai, B., and Stahura, K. A. 2015. "The Marketing Effectiveness of Social Media in the Hotel Industry a Comparison of Facebook and Twitter," Journal of Hospitality & Tourism Research (39:2), pp. 147-169.
Liang, H., Saraf, N., Hu, Q., and Xue, Y. 2007. "Assimilation of Enterprise Systems: The Effect of Institutional Pressures and the Mediating Role of Top Management," MIS quarterly), pp. 59-87.
Lin, R., and Utz, S. 2015. "The Emotional Responses of Browsing Facebook: Happiness, Envy, and the Role of Tie Strength," Computers in human behavior (52), pp. 29-38.
Liu, H. 2007. "Social Network Profiles as Taste Performances," Journal of Computer-Mediated Communication (13:1), pp. 252-275.
Luam, P., Yang, J.-C., and Chiu, Y.-P. 2015. "Why People Check in to Social Network Sites," International Journal of Electronic Commerce (19:4), pp. 21-46.
MacKenzie, S. B., Lutz, R. J., and Belch, G. E. 1986. "The Role of Attitude toward the Ad as a Mediator of Advertising Effectiveness: A Test of Competing Explanations," Journal of marketing research), pp. 130-143.
Mangold, W. G., and Faulds, D. J. 2009. "Social Media: The New Hybrid Element of the Promotion Mix," Business horizons (52:4), pp. 357-365.
83


McCroskey, J. C., Hamilton, P. R., and Weiner, A. N. 1974. "The Effect of Interaction Behavior on Source Credibility, Homophily, and Interpersonal Attraction," Human Communication Research (1:1), pp. 42-52.
Monteith, M. J. 1996. "Contemporary Forms of Prejudice-Related Conflict: In Search of a Nutshell," Personality and Social Psychology Bulletin (22:5), pp. 461-473.
Mooney, C. Z., Duval, R. D., and Duval, R. 1993. Bootstrapping: A Nonparametric Approach to Statistical Inference. Sage.
Mousavizadeh, M., Koohikamali, M., and Salehan, M. 2015. "The Effect of Central and Peripheral Cues on Online Review Helpfulness: A Comparison between Functional and Expressive Products,").
Mudambi, S. M., and Schuff, D. 2010. "What Makes a Helpful Review? A Study of Customer Reviews on Amazon. Com," MIS quarterly (34:1), pp. 185-200.
Mudambi, S. M., Schuff, D., and Zhang, Z. 2014. "Why Aren't the Stars Aligned? An Analysis of Online Review Content and Star Ratings," 2014 47th Hawaii International Conference on System Sciences: IEEE, pp. 3139-3147.
Ngo-Ye, T. L., and Sinha, A. P. 2014. "The Influence of Reviewer Engagement Characteristics on Online Review Helpfulness: A Text Regression Model," Decision Support Systems (61), pp. 47-58.
Nielsen, A. 2012. "Nielson: Global Consumers’ Trust in “Earned” Advertising Grows in Importance," Business Wire).
Nordgren, L. F., Van Harreveld, F., and Van Der Pligt, J. 2006. "Ambivalence, Discomfort, and Motivated Information Processing," Journal of experimental social psychology (42:2), pp. 252-258.
Norman, A. T., and Russell, C. A. 2006. "The Pass-Along Effect: Investigating Word-of-Mouth Effects on Online Survey Procedures," Journal of Computer-Mediated Communication (11:4), pp. 1085-1103.
Okazaki, S. 2009. "Social Influence Model and Electronic Word of Mouth: Pc Versus Mobile Internet," International Journal of Advertising (28:3), pp. 439-472.
Otterbacher, J. 2009. "'Helpfulness' in Online Communities: A Measure of Message Quality," Proceedings of the SIGCHI Conference on Human Factors in Computing Systems: ACM, pp. 955-964.
Pan, L.-Y., and Chiou, J.-S. 2011. "How Much Can You Trust Online Information? Cues for Perceived Trustworthiness of Consumer-Generated Online Information," Journal of Interactive Marketing (25:2), pp. 61-1 A.
84


Pang, B., and Lee, L. 2008. "Opinion Mining and Sentiment Analysis," Foundations and trends in information retrieval (2:1-2), pp. 1-135.
Park, D.-H., Lee, J., and Han, I. 2007. "The Effect of on-Line Consumer Reviews on Consumer Purchasing Intention: The Moderating Role of Involvement," International Journal of Electronic Commerce (11:4), pp. 125-148.
Pavlou, P. A. 2003. "Consumer Acceptance of Electronic Commerce: Integrating Trust and Risk with the Technology Acceptance Model," International journal of electronic commerce (7:3), pp. 101-134.
Pavlou, P. A., and Dimoka, A. 2006. "The Nature and Role of Feedback Text Comments in Online Marketplaces: Implications for Trust Building, Price Premiums, and Seller Differentiation," Information Systems Research (17:4), pp. 392-414.
Pavlou, P. A., and Gefen, D. 2004. "Building Effective Online Marketplaces with Institution-Based Trust," Information systems research (15:1), pp. 37-59.
Petty, R., and Cacioppo, J. T. 2012. Communication and Persuasion: Central and Peripheral Routes to Attitude Change. Springer Science & Business Media.
Petty, R. E., and Cacioppo, J. T. 1986. The Elaboration Likelihood Model of Persuasion. Springer.
Petty, R. E., Priester, J. R., and Brinol, P. 2002. "Mass Media Attitude Change: Implications of the Elaboration Likelihood Model of Persuasion," Media effects: Advances in theory andresearch (2), pp. 155-198.2002.
PurelyBranded.com. 2012. "Why People ‘Like’ and ‘Unlike’ Brands on Facebook."
Qiu, L., Pang, J., and Lim, K. H. 2012. "Effects of Conflicting Aggregated Rating on Ewom Review Credibility and Diagnosticity: The Moderating Role of Review Valence," Decision Support Systems (54:1), pp. 631-643.
Quaschning, S., Pandelaere, M., and Vermeir, I. 2015. "When Consistency Matters: The Effect of Valence Consistency on Review Helpfulness," Journal of Computer-Mediated Communication (20:2), pp. 136-152.
Racherla, P., andFriske, W. 2012. "Perceived ‘Usefulness’ of Online Consumer Reviews: An Exploratory Investigation across Three Services Categories," Electronic Commerce Research and Applications (11:6), pp. 548-559.
Raguseo, E., and Vitari, C. 2017. "The Effect of Brand on the Impact of E-Wom on Hotels’ Financial Performance," International Journal of Electronic Commerce (21:2), pp. 249-269.
Riegner, C. 2007. "Word of Mouth on the Web: The Impact of Web 2.0 on Consumer Purchase Decisions," Journal of advertising research (47:4), pp. 436-447.
85


Riordan, M. A., and Kreuz, R. J. 2010. "Emotion Encoding and Interpretation in Computer-Mediated Communication: Reasons for Use," Computers in human behavior (26:6), pp. 1667-1673.
Salehan, M., and Kim, D. 2014. "Predicting the Performance of Online Consumer Reviews: A Sentiment Mining Approach."
Salehan, M., and Kim, D. J. 2016. "Predicting the Performance of Online Consumer Reviews: A Sentiment Mining Approach to Big Data Analytics," Decision Support Systems (81), pp. 30-40.
Salehan, M., Kim, D. J., and Lee, J.-N. 2015. "Antecedents, Processes and Consequences of Web Assurance Seals: A Meta-Analysis Approach,").
Salton, G., Wong, A., and Yang, C.-S. 1975. "A Vector Space Model for Automatic Indexing," Communications of the ACM (18:11), pp. 613-620.
Schindler, R. M., and Bickart, B. 2012. "Perceived Helpfulness of Online Consumer Reviews: The Role of Message Content and Style," Journal of Consumer Behaviour (11:3), pp. 234-243.
Sen, S., and Lerman, D. 2007. "Why Are You Telling Me This? An Examination into Negative Consumer Reviews on the Web," Journal of interactive marketing (21:4), pp. 76-94.
Shih, H.-p., Lai, K.-h., and Cheng, T. 2013. "Informational and Relational Influences on Electronic Word of Mouth: An Empirical Study of an Online Consumer Discussion Forum," International Journal of Electronic Commerce (17:4), pp. 137-166.
Shimp, T. A. 1981. "Attitude toward the Ad as a Mediator of Consumer Brand Choice," Journal of advertising (10:2), pp. 9-48.
Skowronski, J. J., and Carlston, D. E. 1987. "Social Judgment and Social Memory: The Role of Cue Diagnosticity in Negativity, Positivity, and Extremity Biases," Journal of personality and social psychology (52:4), p. 689.
Soderlund, M., and Rosengren, S. 2007. "Receiving Word-of-Mouth from the Service Customer: An Emotion-Based Effectiveness Assessment," Journal of retailing and consumer services (14:2), pp. 123-136.
Spool, J. M. 2009. "The Magic Behind Amazon’s 2.7 Billion Dollar Question," User Interface Engineering (http://www. uie. com/articles/magicbehindamazon/f).
Storbeck, J., and Clore, G. L. 2008. "Affective Arousal as Information: How Affective Arousal Influences Judgments, Learning, and Memory," Social and Personality Psychology Compass (2:5), pp. 1824-1843.
86


Sun, T., Youn, S., Wu, G., and Kuntaraporn, M. 2006. "Online Word-of-Mouth (or Mouse): An Exploration of Its Antecedents and Consequences," Journal of Computer-Mediated Communication (11:4), pp. 1104-1127.
Sundar, S. S., and Kim, J. 2005. "Interactivity and Persuasion: Influencing Attitudes with Information and Involvement," Journal of interactive advertising (5:2), pp. 5-18.
Sussman, S. W., and Siegal, W. S. 2003. "Informational Influence in Organizations: An Integrated Approach to Knowledge Adoption," Information systems research (14:1), pp. 47-65.
Svensson, A. 2011. "Facebook-the Social Newspaper That Never Sleeps-a Study of Facebook Ewom’s Persuasiveness on the Receivers,").
Taboada, M., Brooke, J., Tofiloski, M., Voll, K., and Stede, M. 2011. "Lexicon-Based Methods for Sentiment Analysis," Computational linguistics (37:2), pp. 267-307.
Tang, C., and Guo, L. 2015. "Digging for Gold with a Simple Tool: Validating Text Mining in Studying Electronic Word-of-Mouth (Ewom) Communication," Marketing Letters (26:1), pp. 67-80.
Van Kleef, G. A. 2010. "The Emerging View of Emotion as Social Information," Social and Personality Psychology Compass (4:5), pp. 331-343.
Vogt, J., De Houwer, J., Koster, E. H., Van Damme, S., and Crombez, G. 2008. "Allocation of Spatial Attention to Emotional Stimuli Depends Upon Arousal and Not Valence," Emotion (8:6), p. 880.
Wang, X.-T. 2006. "Emotions within Reason: Resolving Conflicts in Risk Preference," Cognition and emotion (20:8), pp. 1132-1152.
Wang, Z. 2010. "Anonymity, Social Image, and the Competition for Volunteers: A Case Study of the Online Market for Reviews," The BE Journal of Economic Analysis & Policy (10:1).
Wathen, C. N., and Burkell, J. 2002. "Believe It or Not: Factors Influencing Credibility on the Web," Journal of the American society for information science and technology (53:2), pp. 134-144.
Willemsen, L. M., Neijens, P. C., Bronner, F., and De Ridder, J. A. 2011. "“Highly Recommended!” the Content Characteristics and Perceived Usefulness of Online Consumer Reviews," Journal of Computer-Mediated Communication (17:1), pp. 19-38.
Xu, Q. 2014. "Should I Trust Him? The Effects of Reviewer Profile Characteristics on Ewom Credibility," Computers in Human Behavior (33), pp. 136-144.
87


Yin, D., Bond, S., and Zhang, H. 2014a. "Anxious or Angry? Effects of Discrete Emotions on the Perceived Helpfulness of Online Reviews," Mis Quarterly (38:2), pp. 539-560.
Yin, G., Zhang, Q., and Li, Y. 2014b. "Effects of Emotional Valence and Arousal on Consumer Perceptions of Online Review Helpfulness,").
Zadra, J. R., and Clore, G. L. 2011. "Emotion and Perception: The Role of Affective Information," Wiley interdisciplinary reviews: cognitive science (2:6), pp. 676-685.
Zhang, J. Q., Craciun, G., and Shin, D. 2010. "When Does Electronic Word-of-Mouth Matter? A Study of Consumer Product Reviews," Journal of Business Research (63:12), pp. 1336-1341.
Zhang, W., and Watts, S. 2003. "Knowledge Adoption in Online Communities of Practice," ICIS 2003 Proceedings), p. 9.
Zhang, W., and Watts, S. A. 2008. "Capitalizing on Content: Information Adoption in Two Online Communities," Journal of the Association for Information Systems (9:2), p. 73.
Zhu, L., Yin, G., and He, W. 2014. "Is This Opinion Leader's Review Useful? Peripheral Cues for Online Review Helpfulness," Journal of Electronic Commerce Research (15:4), p. 267.
88


APPENDIX
Appendix A: Summary of Review Helpfulness Studies
Citation Independent Variables Method Sample
Central cues Peripheral Cues
(Baek et al. 2012) Review length, Negative cords Rating inconsistency, Reviewer ranking, Reviewer real name Hierarchical regression Amazon N=75,226
(Mudambi and Schuff 2010) Review depth (word count) Review extremity star rating Tobit Amazon N= 1587
(Kuan et al. 2015) Review Length, Readability, Review valence Rating inconsistency, Reviewer credibility Probit, Heckman selection model Amazon N=89,362
(Salehan and Kim 2014) Title sentiment, Title length, Review sentiment, Review length Review Longevity Negative binomial Amazon N=2616
(Zhu et al. 2014) Review Readability, Review Length Reviewer expertise, Reviewer online attractiveness Negative binomial Yelp N=16,262
(Huang et al. 2015) Word count Reviewer experience, Reviewer impact, Product Rating Tobit Amazon N=2209
(Yin et al. 2014a) Anxiety (anxiety-related words in a review) Anger (anger-related words in a review) Tobit Yahoo N= 7,322
(Mousavizadeh et al. 2015) Utilitarian cues, Hedonic cues, Review Sentiment, Readability Review extremity, Review length, Title sentiment Negative binomial Amazon N=589
89


Full Text

PAGE 1

E WOM COMMUNICATION IN ONLINE CONSUMER REVIEW SITES AND SOCIAL NETWORKING SERVICES by NAVID AGHAKHANI B.S., Shahid Beheshti Universit, 2007 M.S., University of Malaya, 2011 A thesis submitted to Faculty of the Graduate School of the The U niversity of Colorado in partial fulfi l lment o f the requirements for the degree of Doctor of Philosophy Computer Science and Information Systems Program 201 7

PAGE 2

. ii This thesis for the Doctor of Philosophy degree by Navid Aghakhani has been approved for the Computer Science and Information Systems Program by Jahangir Karimi , Chair Dawn Gregg , Advisor Onook Oh Ilkyeun Ra Date: May 1 3 , 201 7

PAGE 3

. iii Aghakhani, Navid (Ph.D., Computer Science and Information Systems) E WOM Communication in Online Consumer Review Sites and Social Networking Services Thesis directed by Professor Dawn Gregg ABSTRACT Electronic Word of Mouth (eWOM) communication has become an important research topic both for marketing and information systems scholars. To explore this topic, this dissertation is comprised of two essays focusing on adoption of eWOM in Social Networking Services (SNS) and credibility of eWOM in online consumer review sites. In the first essay, I investigate the antecedents of adoption of eWOM on Facebook and I suggest two types of eWOM on Facebook. The first type of eWOM is explicit eWOM, which is deliver ed through written text format. The second type of eWOM is implicit eWOM, which is specific to SNS, and delivered through non implicit eWOM, I adopted the theoreti cal lens of Elaboration Likelihood Model (ELM) and Affect As Information Theory. Data for this study is collected through a survey of 2 02 students. In the second essay, I examine drivers of review usefulness in online consumer review sites. Review usefuln ess, as the flip side of review credibility, has been the most important driver of eWOM adoption and consumer purchase decision making in online consumer review sites. For this exploration, I adopted the theoretical lens of ELM and collected 2,634 online c onsumer reviews and the meta data associated with them from Yelp. These two essays contribute to existing literature in eWOM communication in two important ways. The first essay elucidates how the interplay between central and peripheral cues of ELM and Aff ect As Information theory explain the adoption of explicit and implicit eWOM on Facebook. The second essay contributes to ELM by crafting and operationalizing a new variable of review consistency, which measures the consistency between central and peripher al cues of ELM in the online review context. The results

PAGE 4

. iv of these two essays have important managerial implications, particularly, for social media marketing strategy and improving recommender systems in online consumer review sites. The form and content of this abstract are approved. I recommend its publication. Approved: Dawn Gregg

PAGE 5

. v This dissertation is dedicated to my parents for their love, endless support and encouragement. Also, this dissertation is dedicated to my brother and sister, Nima and Anahita, who have been a great source of motivation and inspiration. Finally, this dissertation is dedicated to all those who believe in the richness of learning.

PAGE 6

. vi ACKNOWLEDGMENETS First and foremost, I wish to thank my committee members who were more than generous with their expertise and precious time. Special thanks to Dr. Dawn Gregg, Dr. Jahangir Karimi and Dr. Onook Oh for their countless hours of reflecting, reading, encouraging, and most of all patience throughout the entire process. Also, I would like to thank Dr. Ilkyeun Ra from the computer science department for serving on my committee. I wish to thank my friends, fellow doctoral students, specially Gise lla Bassani, for their help and support in this journey. Finally, I would like to thank my parents for allowing me to realize my own potential. All the support they have provided me over the years was the greatest gift anyone has ever given me . .

PAGE 7

. vii TABLE OF CONTENTS I. DISSERTATION OVERVIEW ................................ ................................ .............................. 1 Adoption of eWOM in SNS ................................ ................................ ................................ .... 2 Credib ility of eWOM in online consumer review sites ................................ ........................... 4 II. A UNIFIED MODEL FOR ADOPTION OF EWOM ON FACEBOOK ............................... 6 Abstract ................................ ................................ ................................ ................................ .... 6 Introduction ................................ ................................ ................................ .............................. 6 Theoretical Fram ework ................................ ................................ ................................ ............ 9 Elaboration Likelihood Model ................................ ................................ ....................... 11 Affect As Information Theory ................................ ................................ ....................... 14 Hypotheses Development ................................ ................................ ................................ ...... 16 Research Methodology ................................ ................................ ................................ .......... 23 Measurem ent of Constructs ................................ ................................ ............................ 23 Data Collection ................................ ................................ ................................ ............... 24 Sample Profile ................................ ................................ ................................ ................ 25 Data Analysis ................................ ................................ ................................ .................. 25 Discussion ................................ ................................ ................................ .............................. 28 Implications ................................ ................................ ................................ ........................... 30 Theoretical Implication ................................ ................................ ................................ .. 30 Practical Implication ................................ ................................ ................................ ....... 33 Limitations and Future Research Direction ................................ ................................ ........... 35 Conclusion ................................ ................................ ................................ ............................. 36

PAGE 8

. viii III. THE EFFECT OF REVIEW CONSISTENCY ON USEFULNESS OF ONLINE REVIEWS: EVIDENCE FROM REVIEWS IN SERVICE CONTEXT .............................. 43 Abstract ................................ ................................ ................................ ................................ .. 43 Introduction ................................ ................................ ................................ ............................ 44 Theoreti cal Foundation ................................ ................................ ................................ .......... 49 Elaboration Likelihood Model ................................ ................................ ....................... 49 ELM and Review Usefulness ................................ ................................ ......................... 50 EL ................................ ................................ ................................ ......... 51 ................................ ................................ ................................ .. 52 Review Consistency ................................ ................................ ................................ ....... 53 Research Methodology ................................ ................................ ................................ .......... 55 Data Col lection ................................ ................................ ................................ ............... 55 Measures ................................ ................................ ................................ ......................... 55 Empirical Models ................................ ................................ ................................ ........... 60 Discussion ................................ ................................ ................................ .............................. 63 Post hoc Analysis ................................ ................................ ................................ ........... 65 Theoretica l and Practical Contributions ................................ ................................ ................ 68 Theoretical Contributions ................................ ................................ ............................... 68 Practical Contributions ................................ ................................ ................................ ... 69 Limitations and Future Research ................................ ................................ ........................... 69 Concluding Remarks ................................ ................................ ................................ ............. 70 REFERENCES ................................ ................................ ................................ ............................. 78 APPENDIX ................................ ................................ ................................ ................................ ... 89

PAGE 9

. ix Appendix A: Summary of Review Helpfulness Studies ................................ ................ 89 Appendix B: Negative Binomial Regression ................................ ................................ . 90

PAGE 10

. x LIST OF TABLES Table II 1. Measurement Items ................................ ................................ ................................ ..... 38 Table II 2. Reliability and Convergent Validity ................................ ................................ ........... 39 Table II 3. Discriminant Validity ................................ ................................ ................................ . 40 Table II 4. Cross Loadings ................................ ................................ ................................ ........... 40 Table II 5. Common Method Bias ................................ ................................ ................................ 41 Table III 1. Descriptive Statistics ................................ ................................ ................................ . 74 Table III 2. Correlation Matrix ................................ ................................ ................................ ..... 75 Table III 3. Categories of Useful Votes ................................ ................................ ........................ 75 Table III 4. Econometric Analyses, Ordered Logit ................................ ................................ ...... 76 Table III 5. Comparing Successive Models ................................ ................................ .................. 77 Table III 6. Post hoc Analysis ................................ ................................ ................................ ...... 77

PAGE 11

. xi LIST OF FIGURES Figure II 1. Research Model ................................ ................................ ................................ ......... 38 Figure II 2. Empirical Results ................................ ................................ ................................ ....... 42 Figure II 3. Social Advertising in Amazon ................................ ................................ ................... 42 Figure III 1. An example of a Review on Yelp ................................ ................................ ............ 72 Figure III 2. Research Model ................................ ................................ ................................ ........ 72 Figure III 3. Supervised Measure of the Review Consistency ................................ ..................... 73 Figure III 4. The Distribution of Rating Based on Number of Reviews ................................ ...... 73

PAGE 12

. 1 CHAPTER I. DISSERTATION OVERVIE W Research on eWOM has attracted scholars from various disciplines, especially marketing and Information Systems. This is mainly eWOM has different definitions, however, in this dissertation I rely on the definition by Henning Thureau et al. (2004) y positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of (Hennig Thurau et al.,2004, p. 39 ) . Studies on impact of eWOM c ommunication is classified into two levels: market level analysis and individual level analysis (Cheung and Thadani, 2012; Lee and Lee, 2009). The focus of market level analysis is on the effect of eWOM on product sales ( Chevalier and Mayzlin 2 006 ; Raguseo and Vitari 2017 ) . At the individual level, however, the focus of research is on personal influence of the eWOM , where, the sender of the message can affect consumer purchase decision ( Cheung et al. 2012 ; Cheung et al. 2009 ) . T his diss ertation falls in the individual level effect of eWOM communication research. The effect of eWOM communication both at market level and individual level has been studied in four main platforms: online discussion forums ( Schindl er and Bickart 2012 ; Zhang and Watts 2008 ) , online consumer review sites ( Awad and Ragowsky 2008 ; Baek et al. 2012 ; Kuan et al . 2015 ) , blogs ( Chu and K amal 2008 ) and online shopping sites ( Gupta and Harris 2010 ; Park et al. 2007 ) . Online consumer review websites are the most common plat forms to study eWOM. Major published studies in this stream tend to investigate the drivers of eWOM credibility ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ; Kuan

PAGE 13

. 2 et al. 2015 ) have already purch ased. Adoption of eWOM in SNS Social Network Services (SNS) have recently been identified as perfect platforms for spreading eWOM and brand related information ( Chu and Kim 2011 ; Fang 2014 ) . Collaborat ive and social characteristics of SNS has enabled their users to show high level of social presence by connecting to other users, exchanging information and opinions ( Kaplan and Haenlein 2010 ) . The desire to establish and ma intain social relationships also enables eWOM behavior among SNS users ( Chu and Kim 2011 ) . By sharing useful products and brand related information users can help their friends (connections) with their purch ase related decisions ( Fang 2014 ) . The increasing use of SNS can also facilitate close relationships among organizations and consumers as a component of an Integrated Marketing Communication (IMC) ( Mangold and Faulds 2009 ) . Users can search for unbiased product information, share their own consumption related advice and display interpersonal comments publicly ( Fang 2014 ; Luarn et al. 2015 ) . The opinion passing behavior that exists in SNS gives an important , yet overlooked, dimension to SNS in comparison to other platforms of eWOM, which merely focus on opinion seeking and opinion giving as the only two ways of eWOM comm unication ( Norman and Russell 2006 ) . The online passing of information enabled by SNS facilitates the flow of information regardless of geographical and temporal constraints ( Chu and Kim 2011 ) . Consumers can quickly exchange brand related information with a few clicks with their connections as well as global audience who share common interests ( Norman and Russell 2006 ) . As a result, SNS as a platform for online branding and

PAGE 14

. 3 advertising has undergone tremendous growth and advertising spending on SNS in the US is expected to reach $2.6 billi on by 2012 ( Chu and Kim 2011 ) . Among existing SNS, Facebook is the most popular, attracting various corporations to open their own fan pages to advertise their brands. As such, Facebook is widely considered a platform of choice for generating and spreading recommendations about specific brands or products and for engaging in eWOM activities (Chu and Kim, 2011). In Facebook , these recommendations have been categorized into explicit and implicit types ( Eber mann et al. 2011 ) . The option to use direct and indirect communication channels in Facebook is similar to the categorization of recommender systems based on explicit and implicit types. In user profiles, information can be provided via status messages o r in pre defined categories such as likes and interests. The main goal of information provided in a ( Liu 2007 ) . Although the major on the o ther users who read it, because it may refer to the products or services the user likes ( Luarn et al. 2015 ) . In other words, while profile information is not directed specifically at other users, it might have a pot ential unintended recommendation effect, which is considered implicit recommendation. Conversely, explicit recommendations are intentionally provided from one SNS user to other users. Such recommendations may be given through direct communication channels such as webmail like messaging within SNS or as a direct response to recommendation requests in status messages ( Ebermann et al. 2011 ) . Given the emergence of SNS as a new platform of eWOM communication and by considering the difference between explicit and implicit eWOM, the first essay of this dissertation will explore the

PAGE 15

. 4 drivers of eWOM adoption through the theoretical lens of Elaboration Likelihood Model and Affect as information theory. Credibilit y of eWOM in online consumer review sites Trust in the online environment has been an important research topic. Especially, in the space of online transaction s information has been shown to be an impor tant driver of the effective marketplace. For instance, regarding eBay , prior literature has focused on the trust issues between buyers and sellers. These trust issues stem from the inherent information asymmetry problem in the online marketplace, where bu yers have more information about sellers, thus increasing the uncertainty in online transaction s . Prior studies show that the implementation of an online feedback system , where consumers can share their opinion about sellers enhance trust in eBay, thereby, increasing the price premium ( Gefen et al. 2008 ; Pavlou 2003 ; Pavlou and Dimoka 2006 ; Pavlou and Gefen 2004 ) . Another aspect of trust in the generated content in online consumer sites. The overwhelming volume of online reviews presents challenges for consumers seeking relevant and trustworthy reviews ( Baek et al. 2012 ; K uan et al. 2015 ; Mudambi and Schuff 2010 ) . To address this problem, online consumer revi ew websites such as Yelp and Amazon have implemented online social voting systems. An online social voting system is a technological component that allows consumers to diagnose the credibility of online reviews by voting for its usefulness or helpfulness ( Baek et al. 2012 ) . Understanding the factors that impact the helpfulness/usefulness of online reviews is an important question for both academics and practitioners because online review qu

PAGE 16

. 5 from a theoretical perspective, review usefulness (helpfulness), as the proxy for review credibility, is known to be the most important driver for consumers to a dopt eWOM ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Other studies have reported that online reviews that are transactions than less useful reviews ( Baek et al. 2012 ; Yin et al. 2014a ) . From a practical perspective, online retailers that display helpful reviews gain a strategic advantage in ( Connors et al. 2011 ) . Reflecting the important role of social voting systems in the online marketplace, Amazon added an extra $2.7 billion to its revenue by implementing its social voting system and asking consumers to rat e the helpfulness of its online reviews ( Spool 2009 ) . To unders tand the phenomena of credibility of online reviews, the second essay of this dissertation focus on the drivers of review usefulness on Yelp. This essay draws on the burgeoning body of research that emphasizes how the content of online reviews affects revi

PAGE 17

. 6 CHAPTER II. A UNIFIED MODEL FOR ADOPTION OF EWOM ON FACEBOOK Abstract Electronic word of mouth (eWOM) is a concept t hat has gained increased attention from both practitioners and academia. Its importance lies in its simplicity and yet its though social network services (SNS) have emerged as a new platform for eWOM communication, less attention has been paid in the literature to the adoption of eWOM in SNS. Using the elaboration likelihood model (ELM) and the affect as information theory, th is study identifies factors that affect the adoption of eWOM on Facebook. We identify product related attributes of a review, source credibility, peer image, and tie strength as theoretically important variables in our study, and we examine their effect on cognitive and affective attitude. We find that the eWOM type (explicit vs. implicit) moderates the effects of cognitive and affective attitude on the adoption of eWOM. We further find that the effect of cognitive attitude on eWOM adoption is higher when the eWOM is explicit, while the effect of affective attitude is higher when eWOM is implicit. Keywords : Electronic Word of Mouth (eWOM), eWOM Adoption, eWOM type, cognitive attitude, affective attitude Introduction Advances in information technology in general and t he emergence of web 2.0 , in particular, have opened new avenues of research through which scholars can study the effect ( Lee and Youn 2009 ) . eWOM has

PAGE 18

. 7 been d customers about a product or company, which is made available to a multitude of people and ( Hennig Thurau et al. 2004 ) . Most of the previous studies on adoption of eWOM were done on conventional platforms for eWOM communication, such as online consumer review websites ( Awad and Ragowsky 2008 ; Cheung et al. 2009 ; Lee and Lee 2009 ; Lee and Youn 2009 ; Park et al. 2007 ) , blogs ( Chu and Kamal 2008 ; Riegner 2007 ) , and online shopping sites ( Gupta and Harris 2010 ; Lee et al. 2008 ; Pan and Chiou 2011 ) . However, t he emergence of social networking services (SNS) has brought significant attention from the commercial marketplace, with global advertisement spending on SNS predicted to exceed $35 billion in 2015, and has opened a new channel for generating and disseminating eWOM ( LePage 2015 ) . Among existing SNS, Facebook is currently the most popular ( eMarketer.com 2014 ) , accounting for 50% of all social referrals and 64% of social revenue ( Cooper 2015 ) . As such, Facebook is generally considered to be the platform of choice for generating, spreading, and encountering eWOM ( Chu and Kim 2011 ) . While c onsiderable attention has been paid to the communication and adoption of eWOM on conventional eWOM platforms, less attention has been paid to the adoption of eWOM on Facebook. This study is particularly motivated by the failure of prior studies to consider the influence of different eWOM types on the drivers of eWOM adoption on Facebook. Past eWOM adoption studies on both Facebook ( Chu and Kim 2011 ; Fang 2014 ) and conventional eWOM platforms ( Cheung et al. 2012 ; Cheung et al. 2009 ) considered only eWOM generated through written text. However, non textual information in a Facebook ( Keenan and Shiri 2009 ; Luar n et al. 2015 ; Okazaki

PAGE 19

. 8 2009 ) er to eWOM that is delivered through non notion that Facebook provides a forum for communicat ion of both explicit and implicit eWOM, and by using other theoretically important variables, three research questions are addressed: (1) How do the cognitive and affective attitudes play a role in the eWOM adoption process? (2) what are important factors that impact the cognitive and affective attitudes for explicit and implicit eWOM adoption? (3) do eWOM type moderates the impact of cognitive and affective attitudes on eWOM adoption . To address the above questions, we consider eWOM through the theoretical lens of the elaboration likelihood model (ELM) ( Petty and Cacioppo 1986 ) and the affect as information theory ( Forgas 1995 ; Forgas and George 2001 ) . We then identify the factors that have a direct impact on cognitive and affective attitudes by investigating the role of product related attributes of a review, source credibilit y, peer image building, and tie strength in our research model. Next, we investigate whether eWOM type moderates the effects of cognitive and affective attitudes on eWOM adoption. This study, therefore, contributes to the existing literature eWOM adoption in several ways. First, it distinguishes between explicit and implicit eWOM by suggesting that implicit eWOM is more unique to Facebook. Second, it integrates ELM and the affect as information theory to explore how the combination of affective and cognitiv e attitudes explain eWOM adoption process. Third, by empirically examining the moderating impact of eWOM type on the effects of cognitive and affective attitudes on the eWOM adoption, this work contributes to the extant literature in eWOM adoption process.

PAGE 20

. 9 Below we provide the theoretical framework for this study followed by hypotheses development. In the methodology section, we describe an empirical study designed to test the proposed model using a sample data from students at two large public universiti es. We then discuss the measurement scales for the constructs in our research model. Next, we present the limitations and conclusions. Theoretical Framework Facebook dif fers from conventional eWOM communication platforms in three important ways. First, consumers use online retailer websites, forums, and blogs specifically for the purpose of seeking and giving opinions about products and services ( Lee and Lee 2009 ; Sun et al. 2006 ) . In other words, the focus of these platforms is on the eWOM itself. In contrast, the primary drivers of SNS use are social presence and interpersonal connections among users ( Fang 2014 ) . Second, the eWOM content o n Facebook is pushed toward users by appearing on their wall or newsfeed , and the source of this eWOM is not anonymous but rather one of the Facebook friends ( Chu and Kim 2011 ; Fang 2014 ; Luarn et al. 2015 ) . Third, prior eWOM communication studies have considered online r eviews (eWOM generated in a written text format) as the main form of eWOM ( Cheung et al. 2012 ; Cheung et al. 2008 ; Cheung et al. 2009 ; Chu and Kim 2011 ; Fang 2014 ) . On Facebook, users can explicitly st ate their opinions about products or services in a written text format , for example, by publish ing a review in the form of a Facebook status update , by posting comments, or by using direct message mechanism. However, Fac e book users can also show their interest in products and services by using non s to share specific information about a brand on their profiles (such as

PAGE 21

. 10 promotions, new products, etc.). Although the main goal of the inf ormation provided in a ( Liu 2007 ) , it can also have the effect of a recommendation for other members ( Ebermann et al. 2011 ) . For instance, location based button allow users to share information on their thus shows his/her consumption behavior and interest in certain products and services, and th is information has the potential to influence the decision making of other SNS users ( Keenan and Shiri 2009 ; Luarn et al. 2015 ; Okazaki 2009 ) . It is now a common practice for businesses to seek to persuad e consumers to decision making process. A 2012 report shows that almost 90% of Facebook users have a friend in their network had also liked it ( PurelyBranded.com 2012 ) . Given the differences between explicit and implicit eWOM, and in view of the social emotional relationships among Facebook users, in this study we use the lens of informational and relational influence s to consider the process of eWOM adoption on Facebook ( Cheung et al. 2009 ; Hennig Thurau et al. 2004 ; Shih et al. 2013 ) . Prior eWOM communication studies have considered the informational influence of eWOM as the effect of the online review ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2008 ; Cheung et al. 2009 ; Kuan et al. 2015 ) . Consistent with these prior studies, one of the paths for eWOM adoption on Facebook is the informational infl uence of informational influences of eWOM, recent studies have also examined the relational

PAGE 22

. 11 influences and social bond characteristics of a platform in predicting eWOM b ehavior ( Fang 2014 ; Lee et al. 2012 ; Shih et al. 2013 ) . This is particularly relevant in the case of Facebook, where sharing of moods, feelings, thoughts, and locations is considered as a form of social interaction; thus, factors related to relational influences (i.e. socio emotional factors) play an importan t role in the adoption of eWOM on Facebook ( Keenan and Shiri 2009 ; Luarn et al. 2015 ; Okazaki 2009 ) advertisement for those products and brands, and numerous studies have identified the affective attitude of consumers toward advertisements as the main driver for adoption of such ads ( Edell and Burke 1984 ; Huang et al. 2013 ; Leung et al. 2015 ; MacKenzie et al. 1986 ; Shimp 1981 ) . Building on these prior studies, in this study we consider cognitive attitude and affective attitude as the two paths for eWOM adoption on Facebook, and we consider the informational and relational (i.e., the socio emotional) influence factors that affect the cognitive and affective paths. Furthermore, we examine the suitability of the ELM and the affect as information theory for explorin g the drivers of cognitive and affective attitude, respectively. Elaboration Likelihood Model The Elaboration Likelihood Model explains that cognitive attitude change among people can be triggered by two routes of influence, the central route and the peripheral route. The primary difference between these routes lies in the amount of thoughtful information processing that is required of the individual subject ( Bhattacherjee and Sanford 2006 ) . The central route requires an individual to think critically about the issues involved in the argument of a message and to inspect the relative facts and relevance of the message before

PAGE 23

. 12 making a conversant judgment about the message. Therefore, p rocessing a message through the central route demands a high cognitive effort. Perception changes that arise from the central route are more stable and long lasting , as they are based on careful consideration of message cues and are therefore more predicti ve of long term behavior. Conversely, the peripheral route involves the processing of heuristic cues ( Angst and Agarwal 2009 ; Bhattacherje e and Sanford 2006 ; Petty and Cacioppo 1986 ) ; it merely requires an processing a message through the peripheral route entails less cognitive effort, and alterations caused by peripheral route influences are less persistent and predictive of long term behavior. Despite the distinctions between these two routes, in practice, people typically evaluate a message throug h a modest level of employing both routes ( Sussman and Siegal 2003 ) . The suitability of the ELM for understanding how people process messages that are intended to be persuasive has been exam ined previously in information systems (IS) research. Bhattacherjee and Sanford (2006) used the ELM to examine the influence processes for information technology acceptance. Sussman and Siegal (2003) employed the ELM in a non experimental setting to study knowledge adoption via electronic mail by consultants at a public accounting firm. Angst and Agarwal (2009) used the theoretical lens of the ELM to study the adoption of electronic health records in an experimental setting. The ELM has also been used in pa st studies as the theoretical lens for eWOM communication and elaboration. For example, the ELM has been used in studies on the effect of negative online consumer ( Lee et al. 2008 ) , the effect of eWOM on ( Gupta and Harris 2010 ) , perceived blogger credibility and the

PAGE 24

. 13 impact of eWOM argument quality on brand attitudes ( Gupta and Harris 2010 ) , and the credibility of eWOM in online consumer review websites ( Cheung et al. 2012 ) . Based on these prior studies, we believe the ELM provides an appropriate theoretical lens to understand the drivers of cognitive attitude change in the context of eWOM adoption on Facebook. Cognitive attitude has been defined as the degree to which an individual develops beliefs relating to the attitude object ( Shih et al. 2013 ) . The notion of the attitude object varies based on the context of the study. For instance, past studies have examined perceived usefulness in the context of the adop tion of information technology ( Adams et al. 1992 ; Davis 1989 ) and trust in the context of the o nline marketplace ( Gefen et al. 2008 ; Pavlou and Gefen 2004 ) . In the context of online consumer reviews, perceived credibility of online reviews has been examined as the main driver of the adoption of eWOM ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ) . In the context of adoption of eWOM on Facebook, however, we believe it is necessary to consider the differences between Facebook and online consumer review sites in conceptualizing the attitude construct. This perspective is also important in sense that the major role of Facebook is hedonic, meaning that people tend to use Facebook for the purpose of social presence in the form of online interpersonal interactions and relationships ( Kaplan and Haenlein 2010 ) . This perspective is also consistent with the study by ( Shih et al. 2013 ) , wherein th ey examined the drivers of eWOM intention in an online forum. They argued that, because the social interaction among members was one of the most important and integral aspects of the forum, the cognitive attitude construct reflected whether the members bel ieved the use of the online forum to access online reviews was wise, beneficial, and valuable. Consistently, we adopt the same conceptualization of the cognitive attitude in this study, and we define cognitive attitude as the degree to which

PAGE 25

. 14 Facebook users believe accessing online reviews through Facebook is wise, beneficial, and valuable. Building on the ELM, we consider the central and peripheral routes as the antecedents of cognitive attitude. In this study, we consider the product related attributes of the eWOM as the factors that relate to the central route of the ELM. Product related attributes of the eWOM can be defined as the degree to which an eWOM posted on Facebook reflects on the attributes of a product. Since source credibility is considered the most important heuristic driver of eWOM adoption in online consumer review platforms ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2 009 ) , therefore, it is considered as one of factors related to the peripheral route of eWOM adoption. Source credibility is defined as the degree to which a recipient of eWOM on Facebook believes the sender of the eWOM (his/her friend) has knowledge, trustworthiness, credibility, and exper tise regarding the product/service. In addition, tie strength and peer image are considered as social relationship variables and part of the peripheral route of eWOM adoption in our model. They have also been studied in prior studies of eWOM communication via SNS ( Casteleyn et al. 2009 ; Chu and Kim 2011 ; Hansen and Lee 2013 ; Lin and Utz 2015 ; Luarn et al. 2015 ; Svensson 2 011 ) . Tie strength is defined as degree to which a Facebook user believes that his/her friends are close to him/her. Peer image is defined as the degree to which a Facebook user believes that his/her peer u ses Facebook to shape an impression of himself/herself. Affect As Information Theory According to the affect as his/her assessment of the consequences of potential actions and decisions ( Zadra and Clore 2011 )

PAGE 26

. 15 object. People also utilize their emotional attitude as a shortcut for evaluating a target in terms of social behavior. This process is very common when quick jud gment or heuristic processing is required ( Forgas 1995 ) . Prior studies in the areas of management and making processes. For instance, the affect as information theory has been utilized to investigate the decision making in an organizational setting ( Forgas and George 2001 ) . It has also been used to examine the impact ( Eroglu et al. 2003 ) . Although affective issues are often overlooked within the IS field ( Yin e t al. 2014a ) , recent studies have begun to incorporate affect in their conceptual framework. This is especially true in the context of eWOM adoption, where the role of affective cues embodied in a ived helpfulness of online reviews ( Salehan et al. 2015 ; Yin et al. 2014a ) . The affect as information theo ry has also been utilized to understand the role of the emotional status of senders and receivers of eWOM in the eWOM adoption process ( Söderlund and Rosengren 2007 ) . In a similar vein, in the context of SNS, ( Fang 2014 ) found that the affective attitude of Facebook users when reading eWOM can positively influence the eWOM adoption on Facebook. Prior studies indicate that the arousal dimension of the affective attitude has a strong influence on information processing by individuals ( Corson and Verrier 2007 ; Vogt et al. 2008 ) . Arousal reflects the degree to which an individual is excited and stimulated. If someone is aroused, then he/she is likely to make a more positive judgment of the target task

PAGE 27

. 16 prompt response ( Storbeck and Clore 2008 ) . The study by Fang (2015) shows that the level of affe ct (arousal) resulting from reading online reviews is one of the drivers of eWOM adoption on Facebook. In addition to reading online reviews (explicit eWOM), Facebook making process ( Keenan and Shiri 2009 ; Luarn et al. 2015 ; Okazaki 2009 ) . Thus, in line with the affect as information theory, the arousal triggered by exploring e implicit eWOM adoption. Given the fact that Facebook contains both explicit and implicit eWOM, the affective attitude is conceptualized as the degree of arousal resulting from exposure to also highlighted the role of social relationship variables as the driver of affective attitude ( Fang 2014 ; Shih et al. 2013 ) , in our research model, the tie strength and peer image building are also included as the drivers of a ffective attitude . Hypotheses Development The effect of content of online reviews on the eWOM adoption has been examined in prior studies in online consumer review platforms ( Ghose and Ipeirotis 2011 ; Qiu et al. 2012 ; Schindler and Bickart 2012 ) . Schindler and Bickart (2012) found that th e amount of product descriptive statements in a review was associated with review usefulness. In a similar vein, Qiu et al. (2012) found that product related attributes of a review enabled consumers to obtain information about characteristics of products a nd thus helped the consumers with their purchase decisions. Conversely, reviews containing non product related information revealed little information about products, and consumers found those reviews to

PAGE 28

. 17 be less diagnostic and less credible. Since the revi ew text plays an important role in the with regard to notion that the explicit eWOM adoption on Facebook is highly associated with the cognitive path, one would expect that product related attributes have a positive impact on the degree to which an individual develops attitudes about eWOM adoption on Facebook. Thus, we propose the following: H 1 : Product related attributes of a review have a positive effect on the cognitive attitude. Online reviews with a high source credibility tend to support eWOM acceptance ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Prior studies based on reviews collected from Amazon have found that source credibility has a positive impact on the helpfulness of online reviews. Forman et al. (2008) reported, for example, that revealing the identity of a reviewer ha d a positive impact on the helpfulness of reviews. More recent studies ( Baek et al. 2012 ; Kuan et al. 2015 ) have demonstrate d that platform based signals of source credibility, such as the top reviewer badge on Amazon, have a positive impact on a review helpfulness. Despite the fact that source credibility in online consumer review websites is only for the eWOM source ( Cheung et al. 2009 ) , it is still an important indicator of information credibility ( Wathen and Burkell 2002 ) . On Facebook, however, because of the rich social interactions and interpersonal relationships among is more likely to be formed based on that redentials ( Chu and Kim 2011 ) . Facebook friends encompass both close ties, such as immediate family, relatives, and friends that meet face to face on a regular basis, as well as more distant acquaintances that

PAGE 29

. 18 may only interact through virtual channels. Becoming friend s on Facebook allows us ers to have access to personal information and content notifications , such as status updates, comments, photographs, visited places, promotio ns, graduations, and so forth. Thus, Facebook enables users to maintain social relationships and to establish trust, which may extend to other contacts as well ( Chu and Kim 2011 ) . It has been found that the eWOM provided by friends is perceived as more credible than that from anonymous or personally unknown sources because these contacts are embedded in the ( Chu and Kim 2011 ) . O n Fac ebook, therefore, people may consider recommendations from friends or classmates as more credible ( Chu and Choi 2011 ) . This makes SNS a vital source of product information for consumers that greatl y enables the communication of eWOM. Therefore, it is no wonder that marketers have invest ed substantial resources in setting up brand profiles on SNS so as to engage consumers with their brand and to spread positive eWOM through SNS members ( Jansen et al. 2009 ) . Thus, because of the importance of source credibility in explicit eWOM adoption , we propose the following : H 2 : Source credibility has a positive effect on the cognitive attitude. Social identity and relationships among users have been identified as the main focus of SNS, particularly Facebook ( Svensson 2011 ) . As s ocial image is a n asset that Facebook users can use to maintain and enhance their status within their network ( Luarn et al . 2015 ) , some users seek to increase their social identity and present an ideal picture of themselves, rather than the reality ( Casteleyn et al. 2009 ) . Image building is thus used by people who aim to publish content that matches the ideal image of themselves that they wish to create ( Kaplan and Haenlein 2010 ) . Presenting an idyllic image of oneself on Facebook can be done by various means . It can happen either explicitly through posting status updates or implicitly by ( Keenan and Shiri 2009 ) . For instance, past studies

PAGE 30

. 19 show that Facebook users share their location information as an indirect way of enhanc ing their self presentation and social image so that they appear more appealing to others in the social network. Thus, it is less likely that people will recommend a product that they believe will damage their social image ( Luarn et al. 2015 ) . We believe the ef fect of peer image building behavior on eWOM adoption on Facebook can be explained based on underlying inferences about the motives of the friend who posted the eWOM. eWOM r motives for recommending a product or service can be classified as either internal (i.e., self serving reasons) or external (i.e., product related reasons) ( Lee and Youn 2009 ) . If motives are related to a product (external), consumers will perceive the review as helpful. On the other hand, if the inferred motives are self serving (internal), consumers will discount the review ( Sen and Lerman 2007 ) . Accordingly, one may recomm end a product or service for internal reasons (i.e., image building) or external reasons (i.e., product related). A qualitative study by ( Svensson 2011 ) shows that eWOM can be ineffective if it is perceived to be communicated for internal reasons. If the eWOM is only to communicate the desired personality of the sender, this reduces the credibility ; consequently , it makes the message less reliable and may even cause it to be rejected. Therefore, image building behavior of a Facebook member reduces his/her cr edibility among his/her peers on Facebook and makes eWOM recommendation less likely to be accepted. attitude about the suitability of Facebook for accessing eWOM may thus be formed in part based on their p In addition, prior studies have also show n that the social attracti veness of individuals make s them more persuasive in general ( McC roskey

PAGE 31

. 20 et al. 1974 ) them more excited to review the eWOM ( Fang 2014 ) . Consistent with these findings , we b elieve that the image building behavior of an individual on Facebook leads other members Check peer image building on Facebook influences the eWOM adoption by having a negative effect on the r cognitive and affective attitudes . Therefore, we propose the following: H 3 : Peer image building has a negative effect on the cognitive attitude. H 4 : Peer image building has a negative effect on affective attitude. is associated with the strength of their ties to one another ( Granovetter 1973 ) . In that sense, social ties can be classified as either strong or weak. Stron g ties constitute stronger and closer ( Brown and Reingen 1987 ) . People have a wide range of social networks among which to search for information, and this includes both strong ties, such as family members and close friends and weak ties, such as acquaintances. However, dynamic information seeking and product referral are more likely to happen among relationships with strong ties ( Brown and Reingen 1987 ) . Similar to the offline environment, in online settings such as SNS, there exist varying degrees of social relationships among members, and these can also be classified a s either strong or weak ( Chu and Kim 2011 ) . The perceived tie strength established via SNS motivates consumers to communicate with one another and to disseminate product related information ( Chu and Kim 2011 ) . Although both strong and weak ties contribute to the propagation of eWOM o n SNS, weak ties exert a wider personal network to external communities. On the oth er hand, strong ties have a more important impact at the individual and small group level; therefore, similar to the offline

PAGE 32

. 21 environment, information seeking and referral behavior is more likely to happen among those with strong ties ( Chu and Kim 2011 ) . In fact, consumers trust online information posted by friends with whom they perceive a strong social relationship more than information posted by those with whom they perceive a weak social relationship ( Pan and Chiou 2011 ) . Thus, Facebook members may perceive Facebook as a valuable platform for accessing eW OM because they can access eWOM from close ties. In addition, the emotions triggered by reading Facebook posts from users with strong ties are higher compared to posts from those with weak ties ( Lin and Utz 2015 ) . Consistently, it can be ex pected that is higher compared to weak ties. Therefore, we propose the following: H 5 : Tie strength has a positive effect on the cognitive attitude. H 6 : Tie strength has a positive effect on affective attitude. Prior studies on information systems reported that cognitive attitude is the main driver for the adoption of information technology ( Angst and Agarwal 2009 ; Bhattacherjee and Sanford 2006 ) . Cognitive attitude is also examined in eWOM adoption studies of online consumer review platforms. These prior studies conceptualized the cognitive attitude in terms eWOM and found that the perceived credibility of eWOM is positively associated with the eWOM adoption . Cognitive attitude is also conceptualized as the degree to which the use of a specific platform for eWOM i s perceived to be beneficial by consumers ( Leung et al. 2015 ; Shih et al. 2013 ) . This specific persp ective is also relevant to Facebook, where the main goal of the platform is to maintain social relationships among its users. If a Facebook user believes that using Facebook to access eWOM is valuable, he/she is more likely to adopt the eWOM. Therefore, we the following hypothesis is proposed next:

PAGE 33

. 22 H 7 : Cognitive attitude has a positive effect on the eWOM adoption. through the lens of their affective attitude change ( Fang 2014 ; Xu 2014 ) . The affective attitude might prompt judgmental responses that may be faster and more consistent across individuals ( Shih et al. 2013 ) . For instance, st udies in marketing and advertising maintain intention to adopt the ads ( Edell and Burke 1984 ; Huang et al. 2013 ; Leung et al. 2015 ; MacKenzie et al. 1986 ; Shimp 1981 ) . On Facebook, we can also consider implicit eWOM advertisement for a through changes in their affective attitude. In addition, a study by Fang (2015) showed that the level of arousal (affective attitude) derived from reading explicit eWOM affects eWOM adop tion. Therefore, by considering the notion that Facebook encompasses both explicit and on Facebook affects their behavioral intention to adopt the eWOM: H 8 : Affective atti tude has a positive effect on the eWOM adoption. The ELM specifies that consumers use a combination of central and peripheral factors in processing information ( Bhattacherjee and Sanford 2006 ; Petty and Cacioppo 1986 ) . Prior studies have maintained that central factors (i.e., factors related to the review text) have a stronger influence in changing effort required to process information embodied by the review text ( Baek et al. 2012 ; Bhattacherjee and Sanford 2006 ) . On the o ther hand, affective attitude change happens quickly and through heuristic factors such as social and emotional responses toward the reviewer ( Fang 2014 ; Xu 2014 ) . Since the eWOM available on Facebook includes a

PAGE 34

. 23 combination of both explicit and implicit eWOM , we believe the eWOM type (explicit vs. implicit) moderates the effect of cognitive and affective attitude on eWOM adoption. The effect of cognitive attitude on eWOM adoption is higher if the eWOM is explicit and the effect of affective attitude on eWOM ado ption is higher if eWOM is implicit. Therefore, we propose the following : H 9a : The effect of cognitive attitude on eWOM adoption is higher for explicit eWOM. H 9b : The effect of affective attitude on eWOM adoption is higher for implicit eWOM. Research Meth odology Measurement of Constructs The scales used to measure the constructs were adopted from the previous literature, with some alterations made to fit the context of th is study. The instrument was pilot tested with a sample comprising two professors, two Ph.D. students, and 50 university students with appropriate knowledge and experience of using SNS to assess the face and content validity of the measures. The purpose of the pilot study was threefold: (1) to assess the internal and external validity of th e scale items; (2) to estimate potential participation rates for the study; and (3) to provide insight into blind spots and oversights that must be addressed in order to execute the research plan. The central route of the ELM in our model consists of the p roduct related attribut e s of the review (PRA). PRA is defined as the degree to which a review reflects the characteristics of the product or service. The scales for measuring this construct were adapted from ( Qiu et al. 2012 ) . The source credibility (SC) represents the peripheral route of the ELM in our model. It is defined as the scales to measure this construct were adapted from multiple previous studies ( Bhattacherjee and Sanford 2006 ; Cheung et al. 2012 ; Cheung et al. 2009 ) . The cognitive attitude (COGA)

PAGE 35

. 24 reflects the attitude of a person about using Facebook to access online reviews. The scales used to measure this construct were adapted from ( Shih et al. 2013 ) . eWOM adoption (EA) is defined as the degree to which an individual on Facebook perceives that Facebook about a product/service was informative for the purchase decision making process. The scales to measure this construct were adapted from ( Cheung et al. 2009 ) . Peer i mage building (IMG) is defined as the degree to which someone perceives that a friend uses Facebook to build his/her social image. The scales for measuring this construct were adapted from ( Luarn et al. 2015 ) . Tie strength (TS) reflec ts degree of closeness with, and perceived importance of , a Facebook friend . The scales to measure this construct were adapted from ( Chu and Kim 2011 ) . Aff ective attitude (AFFA) reflects the individual s feelings about sharing products and brand information through Facebook. The scales used to measure this construct were adapted from ( Fang 2014 ) . The eWOM type was measured by explicit eWOM and implicit eWOM . Data Collection We used a survey to collect data from students at two large public universities in the western United States. Prior studies have show n that students are a good representative for empirical studies on SNS and constitute the major percentage of Facebook users ( Chu and Kim 2011 ; Lenhart 2009 ) . Empirical data for testing of the hypotheses were collected from business major undergraduate students in October 2016. We began by providing the students with a brief description of the survey, without revealing our hypotheses. In the first section, information. Information about their cognitive and affective attitude s was also collected in this section. In the next section, we

PAGE 36

. 25 asked students to recall a friend who ha d engaged in eWOM communication within the past month, and then we asked the respond ents to select one of the following to describe his/her to the product/se Then we asked participants to answer questions pertaining to the product related attributes of the review credibility, tie strength, and image building, Sampl e Profile In all, a total of 202 usable questionnaires were collected. Forty five percent (45%) of the respondents were female and 55 percent were male. Most of the respondents (50%) were within the age range of 18 21 years, followed by 30 percent within the ages of 22 25 years, and 20 percent over the age of 25. Most of the respondents in our sample had between 200 and 300 friends on Facebook. Data Analysis We use d partial least squares (SmartPLS version 3.0 ) to test the measurement model and the structu ral model. Partial least squares (PLS) analysis was chosen over other analytical techniques for two reasons. First, it simultaneously tests both the measurement model and the structural model. Second, it is more appropriate for analyzing moderating effects , because traditional techniques cannot account for measurement error in exogenous constructs ( Chin 1998a ; Chin 1998b ; Chin et al. 2003 ) . Measurement Model Analysis To examine the psychometric properties of the measurement model, this study examined the composite reliability, converge nt validity, and discriminant validity of the

PAGE 37

. 26 loadings. These loadings once deemed consistent with the underlying construct, were used to assess internal consistency and average variance extracted (AVE). Convergent and discriminant validity is adequate for constructs modeled using two or more reflective exceed 0.70 and load more hi ghly on the constructs they are intended to measure ( Chin 1998a ; Chin 1998b ) . Table 2 shows the composite reliability, average varian ce extracted, and benchmark of 0.7 ( Barclay et al. 1995 ; Chin 1998a ; Chin 1998b ) values were above the recommended level of 0.5 ( Chin 1998a ; Chin 1998b ) . Therefore, we square root of the AVE value for each construct exceeds the correlation between that construct and other constructs ( Chin 1998a ; Chin 1998b ; Fornell and Larcker 1981 ) , thus providing evidence for discriminant validity. Commo n Method Bias Because survey methodologies may be subject to common method bias (CMB), we ran a PLS test for CMB using the common factor approach described by Liang et al. ( 2007 ). We created a common method construct having all the items associated with it; we then modeled each of the 30 indicators as a single indicator construct and created paths between them and the common method construct as well as the theoretical constructs. The results showed that load ings on the theoretical constructs were both high and highly significant. Loadings on the common method construct were low and , in al most all cases , non significant. This indicates that CMB is not a problem in this research ( Liang et al. 2007 ) .

PAGE 38

. 27 Structural Model Unlike covariance based SEM, PLS does not provide summary statistics to allow for assess ment of model (R squared) of the endogenous construct the eWOM adoption , and the sign and significance of the path coefficients is typically used to assess model fit. A bootstrapping appr oach was used to produce 500 random samples of the original sample size from the data set by sampling through replacement. This was necessary to obtain estimates of the standard errors for use in testing the statistical significance of the path coefficient s. Such an approach provides valid estimates of the significance of the path coefficients in the PLS models ( Mooney et al. 1993 ) . A summary of the empirical testing and validation of the theorized casual links is given in Figure 2. The product related attributes of the review, sou rce credibility , peer image building, and tie strength were found to explain 42.6 percent of the variance in cognitive attitude. The image building and t ie strength explain ed 14.9 percent of the variance in affective attitude. The cognitive attitude and af fective attitude, in turn, explain ed 36.9 percent of the variance in eWOM adoption . As we predicted , the effect of the product related attributes of the review on cognitive attitude wa s significant ( =0. 167 , P<0.0 5 ) , supporting H 1 . The effect of source credibility on cognitive attitude wa s also significant ( =0. 483 , P<0.01), supporting H 2. We f ou nd that peer image building ha d a negative effect on the cognitive attitude ( = 0. 186 , P<0.01) and affective attitude ( = 0. 342 , P<0.01 ), supporting H 3 and H 4 , respectively . The effect of tie strength on cognitive attitude (H 5 ) was not significant; however, the effect of tie strength on the affective attitude was significant ( = 0. 178 , P<0.0 5 ), supporting H 6 . As we predicted, the effect of cognitive attitude on the eWOM adoption wa s significant ( =0. 334 , P<0.01), supporting H 7 . The effect

PAGE 39

. 28 of affective attitude on the eWOM adoption too was significant ( =0. 254 , P<0.01), supporting H 8 . Consistent with H9 a and H9 b , the eWOM type was found to moderate the effect of cognitive attitude and affective attitude on the eWOM adoption. We also found that, for explicit eWOM, cognitive attitude had a higher effect on the eWOM adoption ( = 0. 30, P<0.01) and, for implicit eWOM, affective attitude had a higher effect on the eWOM adoption ( = 0. 353, P<0.01). Discussion The main purpose of this study is to move beyond the conventional definition of eWOM as a form of reviews delivered through written text (i.e., explicit eWOM) by highlighting the notion that non Facebo ok can be considered as a new type of eWOM (i.e., implicit eWOM). Against this background, we consider Facebook as a n eWOM communication platform that encompasses both explicit and implicit eWOM, and by building on the ELM and the affect as information the ory, we develop a unified theoretical model that examines the role of eWOM type in the eWOM adoption process on Facebook. The findings of this study are threefold . First, as with studies of eWOM adoption in online consumer review platforms and based on t he theoretical lens of the ELM, we find support for the positive effect of cognitive attitude on eWOM adoption. In addition, based on the affect as information theory, we examined the explanatory power of the a ffective path in eWOM adoption by finding supp ort for the positive effect of affective attitude on eWOM adoption. Reflecting on affective attitude, this study confirms the findings of Fang (2015) concerning the effects of the arousal dimension of affective attitude on eWOM adoption, and it extends the notion of arousal to the emotional outcome of exposure to both explicit and

PAGE 40

. 29 implicit eWOM . Second, regarding the role of eWOM type in the eWOM adoption process, we find that eWOM type moderates the effect of cognitive and affective attitudes on the eWOM a doption. Our results show that, when the eWOM is explicit, as it is in online consumer review platforms, the cognitive attitude has a higher effect on the eWOM adoption . However, if the eWOM type is implicit, the affective attitude has a higher effect on t he eWOM adoption . Third, regarding the drivers of cognitive and affective attitude, our results antecedents of cognitive attitude. Our results show that, when it comes to exp licit eWOM, Facebook users pay attention to product related attributes of the explicit eWOM, and the higher elaboration of those attributes resulted in a higher cognitive attitude. In addition, our results support prior eWOM adoption studies that highlight source credibility as the most important peripheral factor for explicit eWOM adoption. Our model considered the socio interpersonal relationships among Facebook members and examined the effects of tie strength and peer fective attitudes. We did not find support for a positive effect of tie strength on the cognitive attitude. One possible explanation for this result might be the finding from prior studies that consumers accept eWOM from their close ties because they perce ive them as a credible source of eWOM. Thus, tie strength may have an indirect effect on cognitive attitude by influencing perceived source credibility. We further checked this path in our structural model and found that tie strength had a positive effect on the source credibility ( = 0. 44, P<0.01). However, we also found a positive effect of tie strength on the affective attitude, providing support for the role of tie strength as one of the drivers of affective attitude. Our results also highlight the ro le of image building as one of less studied personal conditions that motivates eWOM behavior in SNS, by providing empirical support

PAGE 41

. 30 for image building having a negative effect on eWOM adoption by having a negative influence on both the cognitive and affect ive attitudes. Implications Theoretical Implication This study makes important theoretical contributions to the extant body of literature on eWOM communication. The question of how and under what circumstances consumers adopt eWOM has become an important research topic. Most previous eWOM studies have examine d the eWOM adoption from online consumer review sites ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2008 ; Cheung et al. 2009 ; Mudambi and Schuff 2010 ) . These studies consider ed the perceived credi bility of online reviews as the most important driver of eWOM adoption ( Baek et al. 2012 ; Wang 2010 ) . Other studies, in a similar vein, collect ed online reviews to examine how the attributes of a review text and the characteristics o f the eWOM source influence the perceived helpfulness of online reviews , as the flip side of review credibility ( Cheung et al. 2012 ; Salehan et al. 2015 ; Yin et al. 2014a ) . This study contributes to this body of literature by examining the eWOM adoption in social networking services, and in particular , Facebook. Despite the emergence of SNS as a new platform for eWO M communication, the eWOM adoption o n SNS has been largely neglected in previous studies. In particular, the eWOM adoption o n Facebook is important because of the differences and commonalities between Facebook and the conventional platforms for eWOM commun ication. These differences stem from the notion that social presence and interpersonal relationships between members are the primary reasons for the use of Facebook ( Fang 2014 ; Kaplan and Haenlein 2010 ) ; in other words , t he communication of eWOM itself is not the main emphasis of Facebook. A second important distinction is that, in addition to

PAGE 42

. 31 explicit eWOM , there is evidence that non textual signals such as information shared through e potential to influence purchase decision making of Facebook users ( Keenan and Shiri 2009 ; Luarn et al. 2015 ; Okazaki 2009 ) . We refer to these non textual forms of eWOM here as implicit eWOM. Reflecting on the contextual differences between Facebook and conven tional eWOM communication platforms, and by considering eWOM type, this study contributes to the theoretical underpinning on eWOM adoption studies in the following ways. First, this study contributes to the conceptualization of cognitive and affective atti tudes as the two antecedents of eWOM adoption. Prior eWOM adoption studies of both online consumer review platforms ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ) and Facebook ( Fang 2014 ) have examine d the perceived credibility of the eWOM as the focal cognitive based construct in their theoretical model ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ) . However, we find that perception of the suitability of Facebook for accessing online reviews is an important cognitive based construct for examining the adoption of eWOM o n Facebook. Thus, our findings confirm prior studies ( Leung et al. 2015 ; Shih et al. 2013 ) in takin g into account the usability of the platform for eWOM communications when the social relationships among members of the platform is an integral part of the platform making has been examined in p ast studies ( Eroglu et al. 2003 ; Forgas and George 2001 ; Söderlund and Rosengren 2007 ) , few studies have examined the affective attitude as one of the antecedents of eWOM behavior on Facebook ( Fang 2014 ; Shih et al. 2013 ; Xu 2014 ) . This study contributes to this stream of research by finding evidence that the arousal dimension of the affective attitude affects eWOM adoption. In particul ar, this study extends

PAGE 43

. 32 the findings from Fang (2015) on the effect of arousal on eWOM adoption by providing evidence that the arousal derived from exposure to both explicit and implicit eWOM can serve as a source of information and thus can affect eWOM ado ption. Second, this study confirms the suitability of the ELM and the affect as information theory to understand the drivers of the cognitive and affective attitude, thus providing empirical support for the integration of these two theories to understand e WOM adoption on Facebook. Similar to previous studies of online consumer review platforms ( Ghose and Ipeirotis 2011 ; Qiu et al . 2012 ; Schindler and Bickart 2012 ) , our results show that product related attributes of a review on Facebook have a positive effect on the cognitive attitude. Our results also confirm prior studies that identify source credibility as the most important of ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ) . In addition, this study examined tie strength and peer image as two social interpersonal peripheral factors that affect both cognitive and affective attitudes . While prior studies exa mined tie strength as one of drivers of eWOM engagement as well as information seeking and giving behavior in SNS ( Chu and Kim 2011 ; Hansen and Lee 2013 ; Pan and Chiou 2011 ) , our results show that tie strength did not have a direct effect on the cognitive attitude. However, further analysis showed that it had an indirect ef fect on cognitive attitude by having a positive effect on perceived source credibility. Our results also show that peer image building has a negative effect on cognitive attitude, thus contributing to prior literature ( Folkes 1988 ; Sen and Lerman 2007 ) that examined the role of internal (self serving) and external (product related) motives in adopting product recommendations. Our results show that image building behavior among Facebook users can be considered as an internal motive by Facebook members an d that Facebook users tend to discount eWOM

PAGE 44

. 33 generated by users who use published content such as product and brand information to promote their social image. Third, this study examined the role of eWOM type in the eWOM adoption process by finding that the eWOM type moderates the effect of the cognitive and affective attitudes on the eWOM adoption process. Our results show that the cognitive attitude has a higher effect on eWOM adoption when the eWOM is explicit. Conversely, the affective attitude has a hig her effect on eWOM adoption when the eWOM is implicit. Practical Implication The results of this study also offer two important practical contributions. First, our results show that , as with online consumer review platforms, the content of Facebook review s play s decision making process. Given Facebook role as a platform where the main purpose is hedonic rather than for communication of eWOM, the perception of Facebook users about suitability for accessing online reviews is particularly important for eWOM adoption from this platform . Our results show that, when it comes to explicit eWOM, Facebook does not differ significantly from online consumer review platforms, and the extent to which the eWOM elaborates on the attributes suitability of Facebook for accessing online reviews. This finding may hold important implications for the current practice of social advertisement in online shopping platform s. Many online shopping platforms allow buyers to share their purchase information on an SNS immediately after their purchase. For i nstance, Figure 3 1 shows an example of an Amazon purchase, where the buyer can share information about his/her purchase thro ugh various channels , including Facebook. 1 Source: https://vwo.com/blog/high converting thank you pages/

PAGE 45

. 34 The content and metadata for this purchase are pre filled with generic content. As our result shows that the product related attributes of the eWOM are positively associated with eWOM adoption, our suggestion for online retailers and shopping platforms would be to persuade Facebook users to modify the content of such purchase information based on the attributes of the product that they like best. Second , our results show that the affective attitude of consumers toward eWOM o n Facebook is one of the drivers of eWOM adoption o n Facebook and that the affective attitude has a higher effect on eWOM adoption when the eWOM type is implicit. In addition, our res ults show that tie strength and peer image are two social interpersonal relationship variables that influence affective attitude. These findings have important implications for the practice of sponsored story advertising o n Facebook. Such sponsored ads are commonly accompanied by social cues such as businesses/products or brands that other Facebook friends have either hecked in or iked. A prior study using a randomized field experiment show ed that the presence of close ties as a social signal increase d the likelihood of response to such ad vertisement s ( Bakshy et al. 2012 ) . While our study confirms this finding, the notion of image building behavior among Facebook users demands further attention for the practice of sponsored advertisement. While further investigation would be needed in order to quantify the measure of image building, we believe that one possible approach to mitigate the negative effect of image building on eWOM adoption is to consider ikes, C heck ins and other socio economic signals before including the m as a social component of a sponsored advertisement. For instance, if a Facebook user new brand page deviates from his/her past concerning the same product or service type, this new can be excluded from the sponsored

PAGE 46

. 35 advertisement. Alternatively, if multiple Facebook friends the same product/brand page, because of the positive effect of a number of social cues on the likelihood of response to an advertisement ( Bakshy et al. 2012 ) , t hese Facebook members can be included in the ad vertisement but not as the main social cue. Limitations and Future Research Direction One of the major limitations of this study is t he use of university student s as the sample respondent s, which may raise questions about the generalizability of the results. Although previous studies have shown that students are a good representative of the population for empirical studies on SNS, and t hat they are known to constitute the major percentage of Facebook users ( Chu and Kim 2011 ; Lenhart 2009 ) , our findings should be generalized only with caution. Our sample also lacks cultural and language diversity, which may limit its general application to other cultures. Language, as the vehicle for a textual message, has a significant effect on the perception of the receiver ; h ence, the adoption of eWOM may be different across diverse languages. Also, a study by ( Chu and Choi 2011 ) showed that differences in culture influence the acceptance of eWOM. People from collectivist cultures engage in greater levels of information seeking and information giving behaviors on SNS than their individualistic counterparts. Moreover, there are differences between collectivist and individualistic societies in terms of network structure. People from collect ivist cultures have more strong ties on SNS, while people from individualistic cultures tend to have weaker ties. Finally, people from collectivist cultures have higher levels of trust in their SNS contacts, and they are thus more likely to be influenced b y eWOM than their individualistic counterparts. Hence, in the future, comparative studies on other social networking services in different cultures would be fruitful. Furthermore, we use self reported

PAGE 47

. 36 data in this study, which is subject to social desirabi lity response bias ( Arnold and Feldman 1981 ) . To address this bias, it may be useful for future research to analyze textual information collected from SNS and to look at how different characteristics of the text, such as reada bility and descriptiveness, influence its acceptance. Future research may also conduct experiments to investigate the acceptance of eWOM. Previous research has shown that computer mediated communications (CMC) can effectively transfer emotions. Moreover, t he emotions contained in a message transferred through CMC significantly influence how the message is processed and interpreted by its recipient ( Berger and Milkman 2012 ; Riordan and Kreuz 2010 ) . Hence, future research may consider the effect of message sentiment on the image building behavior on eWOM adoption, quantifying this construct based on past social behavior would be fruitful for social media marketers . C onclusion Although online consumer review websites are the major platform for eWOM communication, the emergence of SNS has created a new avenue for both consumers and marketers to review and share brand/product related information. To th at end, this study highlights the differences between SNS and conventional platforms for eWOM communication. We argue that, while explicit eWOM is common across all platforms for eWOM communication, implicit eWOM is only salient on SNS. This study draws on the ELM and the af fect as information theory to understand the drivers of cognitive and affective attitudes, thus contributing to prior eWOM adoption studies by providing empirical evidence of the benefits of integrating these two theories to explore the eWOM adoption proce ss in the presence of both explicit and implicit eWOM. Our results show that product related

PAGE 48

. 37 attributes of a review, positively affect the cognitive attitude. In addition, we find that source credibility also has a positive effect on the cognitive attitude. Furthermore, our study considers tie strength and peer image as two social interpersonal relationship variables that influence both cognitive and affective attitude. Our results show that image building has a negative effect on both the cognitive and affective attitudes and that tie strength influences eWOM adoption by having a positive effect on the affective attitude; however, we did not find support for the positive effect of tie strength on the cognitive attitude. Our study shows that the eWOM type moderates the effect of cognitive and affective attitudes on the adoption of eWOM. The effect of cognitive attitude on the eWOM adoption is higher when the eWOM is explicit. Conversely, the effect of affective attitude on the eWOM adoption is higher when eWOM is implicit . The results of this study contribute to the extant body of eWOM adoption literature by including eWOM type in our theoretical model of eWOM adoption and by investigating its moderating impact on the effects of cognitive and affective attitudes on eWOM adoption.

PAGE 49

. 38 TABLES AND FIGURES Figure II 1 . Research Model Table II 1 . Measurement Items Construct Items Reference Product related attributions of the review (PRA) PRA 1 .The content of his/her review was based on the product/service PRA 2 . His/her review reflects the characteristics of the product/service ( Qiu et al. 2012 ) Source credibility (SC) SC 1 . He/she was knowledgeable about the product/service SC 2 . He/she was trustworthy SC 3 . He/she was credible SC 4 . He/she appears to be an expert on the product/service ( Bhattacherjee and Sanford 2006 ; Cheung et al. 2012 ; Cheung et al. 2009 ) Cognitive attitude (COGA) COGA 1 . Using Facebook to access reviews about products/services is wise COGA 2 . Using Facebook to access reviews about products/services is beneficial COGA 3 . Using Facebook to access reviews about products/services is valuable ( Shih et al. 2013 ) eWOM adoption (EEA) EA 1 . post on Facebook contributed to my knowledge of the product/service ( Cheung et al. 2009 )

PAGE 50

. 39 Construct Items Reference EA 2 . product/service made it easier for me to make a purchase decision (e.g., to purchase or not to purchase) EA 3 . product/service has enhanced my effectiveness in making a purchase decision EA 4 . product/service motivated me to make a purchase decision. Peer i mage building ( P IMG) P IMG 1 . He/she uses Facebook to shape an impression of himself/herself P IMG 2 . He/she uses Facebook to build image of himself/herself ( Luarn et al. 2015 ) Tie strength (TS) TS 1 . He/she is important to me TS 2 . He/she is close to me TS 3 . I contact him/her frequently ( Chu and Kim 2011 ) Affective attitude (AFFA) When I see friends share information about products/services on Facebook : AFFA 1 AFFA 2. frenzied AFFA 3. ( Fang 2014 ) Table II 2 . Reliability and Convergent Validity Average Variance Extracted (AVE) Composite Reliability Cronbach's Alpha EA 0.712 0.907 0.862 AFFA 0.829 0.936 0.897 COGA 0.932 0.976 0.964 P IMG 0.94 0.969 0.936 PRA 0.874 0.932 0.862 SC 0.802 0.942 0.917 TS 0.888 0.96 0.937

PAGE 51

. 40 Table II 3 . Discriminant Validity EA AFFA COGA PIMG PRA SC TS EA 0.844 AFFA 0.284 0.911 COGA 0.388 0.403 0.966 P IMG 0.329 0.343 0.344 0.969 PRA 0.591 0.162 0.46 0.197 0.935 SC 0.527 0.346 0.608 0.258 0.554 0.896 TS 0.346 0.18 0.234 0.008 0.333 0.44 0.942 Note : Diagonal values are square roots of AVEs. Table II 4 . Cross Loadings Affective Cognitive Image Tie Adoption SC PRA AATT 1 0.902 0.372 0.325 0.17 0.303 0.339 0.173 AATT 2 0.927 0.357 0.326 0.172 0.227 0.29 0.096 AATT 3 0.903 0.372 0.283 0.149 0.241 0.312 0.173 CATT 1 0.408 0.959 0.328 0.232 0.344 0.584 0.417 CATT 2 0.38 0.968 0.321 0.229 0.362 0.588 0.453 CATT 3 0.381 0.969 0.346 0.218 0.415 0.589 0.461 PIMG 1 0.346 0.339 0.971 0.027 0.312 0.246 0.213 PIMG 2 0.318 0.327 0.968 0.012 0.328 0.256 0.168 TS 1 0.177 0.238 0.004 0.941 0.269 0.428 0.333 TS 2 0.213 0.219 0.018 0.966 0.342 0.425 0.297 TS 3 0.104 0.201 0 0.92 0.382 0.385 0.313 EA 1 0.227 0.355 0.36 0.253 0.81 0.471 0.669 EA 2 0.242 0.419 0.341 0.31 0.918 0.525 0.583 EA 3 0.269 0.335 0.257 0.343 0.91 0.482 0.467 EA 4 0.224 0.16 0.122 0.258 0.721 0.263 0.226 SC 1 0.292 0.607 0.252 0.382 0.526 0.892 0.54 SC 2 0.321 0.537 0.205 0.45 0.45 0.931 0.51 SC 3 0.296 0.522 0.212 0.414 0.441 0.92 0.498 SC 4 0.332 0.501 0.254 0.328 0.463 0.836 0.428 PRA 1 0.118 0.323 0.175 0.276 0.507 0.445 0.906 PRA 2 0.174 0.503 0.192 0.337 0.587 0.57 0.962

PAGE 52

. 41 Table II 5 . Common Method Bias Indicators Theoretical Construct Loading T stat P value Common Method Factor Loading T stat P value AATT 1 0.861 25.553 p<0.01 0.053 1.158 p>0.05 AATT 2 0.954 46.329 p<0.01 0.043 1.314 p>0.05 AATT 3 0.917 36.941 p<0.01 0.008 0.191 p>0.05 CATT 1 0.973 45.263 p<0.01 0.017 0.707 p>0.05 CATT 2 0.976 54.003 p<0.01 0.011 0.425 p>0.05 CATT 3 0.948 50.232 p<0.01 0.027 1.206 p>0.05 EA 1 0.681 8.5 p<0.01 0.161 1.876 p>0.05 EA 2 0.828 16.811 p<0.01 0.107 1.825 p>0.05 EA 3 0.938 28.29 p<0.01 0.028 0.618 p>0.05 EA 1 0.939 10.789 p<0.01 0.272 2.782 p<0.01 PIMG 1 0.966 90.527 p<0.01 0.008 0.456 p>0.05 PIMG 2 0.973 96.396 p<0.01 0.008 0.456 p>0.05 PRA 1 1.025 41.234 p<0.01 0.126 3.431 p<0.01 PRA 2 0.85 23.153 p<0.01 0.126 3.431 p<0.01 SC 1 0.736 11.682 p<0.01 0.168 2.628 p<0.01 SC 2 0.992 19.627 p<0.01 0.064 1.101 p>0.05 SC 3 1.024 19.192 p<0.01 0.112 1.686 p>0.05 SC 4 0.818 9.969 p<0.01 0.019 0.202 p>0.05 TS 1 0.93 39.519 p<0.01 0.005 0.149 p>0.05 TS 2 0.96 56.015 p<0.01 0.007 0.23 p>0.05 TS 3 0.939 43.716 p<0.01 0.012 0.322 p>0.05

PAGE 53

. 42 Figure II 2 . Empirical Results Figure II 3 . Social Advertising in Amazon

PAGE 54

. 43 CHAPTER III. THE EFFECT OF REVIEW CONSISTENCY ON USEF ULNESS OF ONLINE REVIEWS: EVIDENCE FR OM REVIEWS IN SERVIC E CONTEXT Abstract Online reviews pl ay a significant role in product diagnostics as well as in the online retail websites such as amazon.com and online consumer review websites such as yelp.com has ena bled users to share their experiences with other consumers seeking trustworthy information about specific products and services. Further, the social voting features in Yelp and Amazon are examples of mechanisms that augment trust by allowing consumers to v ote for the usefulness of online reviews. Employing the Elaboration Likelihood Model (ELM), this study analyzes Yelp data to better understand the antecedents of review usefulness. The data analyses show that unsupervised and supervised measures of review consistency, representing both the central and peripheral cues of ELM, have positive effects on review usefulness. The analyses also show the negative effect of review rating and the positive effect of source credibility, reflecting the non trivial effect of the peripheral cues of ELM. These findings provide new insights on how sentiment mining and machine learning techniques can be utilized to analyze the quality of online reviews. Our findings also provide insights for online retailers on how to manage an d sort reviews in their websites. Keywords: online review, review consistency, Yelp, social voting, e Word of Mouth, machine learning, text mining, sentiment analysis.

PAGE 55

. 44 Introduction Advances in information technology and the emergence of Web 2.0 have enabled consumers to freely share their opinions and experiences about products and services anytime, anywhere. Like other forms of electronic word of mouth (eWOM) , online reviews play an inescapable role in reducing the uncertainty in online shopping by g uiding consumers in their purchase decision making process ( Kuan et al. 2015 ; Yin et al. 2014a ) . It has been reported that 92 percent of consumers read online revie ws prior to purchases and 89 percent of consumers make decision in their purchase based on online reviews ( eMarketer 2010 ) . Online retailer websites such as Amazon and online consumer review websites such as Yelp are among the most used product recommendation platforms for online consumers. According to a recent report ( Nielsen 2012 ) , after family members and friends, consumer review websites are rated as the most trustworthy and influential source of information on products and services. This means that people view online reviews as more reliable and less biased than any other information available on products and services ( Lee and Youn 2009 ) . However, t he overwhelming volume of online reviews presents challenges for consumers seeking relevant and trustworthy reviews ( Baek et al. 2012 ; Kuan et al. 2015 ; Mudambi and Schuff 2010 ) . To address this problem, online consumer review websites such as Yelp and Amazon have implemented online social voting systems. An online soc ial voting system is a technological component that allows consumers to diagnose the credibility of online reviews by voting for its usefulness or helpfulness ( Baek et al. 2012 ) . Understan ding the factors that impact the helpfulness/usefulness of online reviews is an important question for both academics and practitioners because online review quality is directly related to rom a theoretical

PAGE 56

. 45 perspective, review usefulness (helpfulness), as the proxy for review credibility, is known to be the most important driver for consumers to adopt eWOM (electronic word of mouth) ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Other studies have reported that online reviews that are perceive online transactions than less useful reviews ( Baek et al. 2012 ; Yin et al. 2014a ) . From a practical perspective, online retailers that display help ful reviews gain a strategic advantage ( Connors et al. 2011 ; Yin et al. 2014a ) . Reflecting the important role of social voting systems in the online marketplace, Amazon added an extra $2.7 billion to its revenue by implementing its social voting system and asking consumers to rate the helpfulness of its online reviews ( Spool 2009 ) . The extant body of research on online reviews has mainly focuse d on identifying the factors that affect the usefulness (in Yelp) or helpfulness (in Amazon) of online reviews ( Baek et al. 2012 ; Cao et al. 2011 ; Forman et al. 2008 ; Kuan et al. 2015 ; Mudambi and Schuff 2010 ; Salehan and Kim 2014 ; Salehan and Kim 2016 ; Zhang et al. 2010 ) . They have examined the relationship between characteristics of online reviewers and review quality, or the relationship between th e number of star rating and review quality, where the review quality is operationalized as the number of useful votes (in Yelp) or helpful votes (in Amazon) received from other online users ( Forman et al. 2008 ; Mudambi and Schuff 2010 ) . More recent studies have begun to use text mining and sentiment mining methods to investigate how the online review content impacts review usefulness ( Baek et al. 2012 ; Kuan et al. 2015 ; Salehan et al. 2015 ) . These studies focus on the effect of the valence of online reviews (i.e., negative, neutral, and positive) or the emotion bearing words in the review on review helpfulness.

PAGE 57

. 46 These prior studie s focus either on non verbal features of review rating or verbal feature of review text, failing to consider that both review text and review rating are inseparably entangled ( Tang and Guo 2015 ) . In practice, consumers tend to simultaneously consider both non verbal star rating and verbal review text to assess the usefulness of an online review, t he refore, inconsistency between a review text and the accompanying numerical rating increases the uncertainty and decreases the value of online reviews ( Mudambi et al. 2 014 ) . In addition, p rior studies reported the biases associated with using single factor measure (i.e. numerical star rating) in evaluating online reviews. 2 experience on Yelp: stopped in for a quick bite to eat at Dojo, a restaurant in New course , I did I ate there and experienced it for myself. The food was okay. The service was okay. On average, it was average. So I went to rate the restaurant on Yelp with a strong idea of the star rating I would give it. I logged in, navigated to the page and clicked the button to write the review. I saw that, immediately to the right of right under her bright red five star rating. three, but Shar had a point: Her review moved me. And I gave the place a four. As it turns out, my behavior is not uncommon. In fact, this type of social influence is dramatically biasing online ratings one of the most trusted sources o f consumer confidence in e As it can be seen, Sinan Aral attributes the bias in numerical star rating to positive social influence that exists in e commerce websites. Other studies also maintain that a single factor measure (i.e., star rating) is not a sufficient proxy for review quality. For example, the J 2 http://sloanreview.mit.edu/article/the problem with online ratings 2/

PAGE 58

. 47 question the reliability of online review system itself ( Gao et al. 2015a ; Hu et al. 2009 ) . Hu et al. (2009) point out that people tend to post online reviews when they are extremely happy or extremely unhappy. That means people who feel an average level of satisfaction with a should not solely rely on the simple average [star should incorporate other aspects to make better decisions about the usefulness of a specific online review ( Hu et al. 2009 ) . Reflecting on these prior studies, we theorize the consistency of online reviews , meaning the level of consistency between non verbal star rating and verbal review text . By drawing upon Elaboration Likelihood Model ( Petty and Cacioppo 1986 ) , we explore the votes received by online reviewers. In particular, to address the inconsistency problem in online reviews, we craft and operationalize a review consistency variable, which is conceptualized as a two factor measure and operationalized as the degree of consistency between the review text (i.e., the qualitative assessment o f the online review text, as a central cue) and the numerical star rating (i.e., the quantitative rating of the online review, as a peripheral cue). By using the two factor measure (i.e., the review consistency variable) and other theoretically important v ariables , we explore the factors that are important in identifying useful and trustworthy online reviews. For this exploration, we collect and analyze 2,634 restaurant reviews from Yelp. We use Yelp as the context of this study because consumers rely heav ily on online reviews about services due to higher uncertainty associated with services compared to products ( Racherla and Friske 2012 ) . In addition, prior r esearch shows that the inconsistency in online reviews is

PAGE 59

. 48 more prevalent in experience goods (e.g., food experiences in restaurants), compared to search goods ( Mudambi et al. 2014 ) . To empirically test our proposed model (refer to Figure 3), we employ sentiment mining and machine learning techniques i n combination with ordered logit and negative binomial regression. The data analyses show that unsupervised and supervised measures of review consistency have positive effects on review usefulness. In addition, the new two factor variable of review consist ency , which we developed for this study, predicts the usefulness of online reviews above and beyond prior single factor variables . This study makes important contributions to both theory and practice. In theory, it advances our knowledge of the factors th at are important to identify useful and trustworthy online reviews. In particular, the review consistency variable significantly advances ELM by presenting empirical evidence that an online review system becomes more useful 3 and trustworthy when the non ve rbal star rating (i.e., a peripheral cue) is complemented with the verbal content of online review (i.e., a central cue). In that regard, this study also contributes to the existing literature on online trust but in the new context of online reviews. In pr actice, this study makes an important contribution, as online review platform providers may improve their sorting algorithms by incorporating the proposed review consistency variable into their recommendation systems. The remainder of this paper is struct ured as follows. The next section discusses the suitability of ELM as the theoretical premise of this study. The research model and related hypotheses are presented in this section. In the research methodology section, we discuss the 3 Since review usefulness and review helpfulness are the same mechanism implemented in social voting system of Yelp and Amazon respectively, from now on in this study we just use the term usefulness for the sake of consistency and avoiding any confusions.

PAGE 60

. 49 data, measurements, an d econometric model used in this study. Methods of developing measurements for review consistency are also discussed in detail. In the discussion section, the results of our data analysis are reported. Finally, theoretical and practical implications of thi s study are discussed, followed by a treatment of the limitations of this research and future directions for study. Theoretical Foundation Elaboration Likelihood Model According to Elaboration Likelihood Model (ELM) ( Petty and Cacioppo 1986 ) , the persuasiveness of a message can be determined by the central and peripheral cues. The central cues r equire an individual to think critically about the issues involved in the argument of the verbal message and to inspect the relative facts and relevance of the message before making a conversant judgment about the message. Therefore, processing a message t hrough the central cues demands high cognitive effort. Conversely, processing a message through the peripheral cues involves less cognitive effort , because the influence of a message is highly associated with heuristic cues (Angst and Agarwal 2009). ELM po sits that central and peripheral cues affect the judgment of the message recipients to be either persuaded or not persuaded by the received message . The central cues have a persistent attitude, which in turn, influences the receive other hand, the peripheral cues have a temporal effect on the attitude of a receiver of the message, which may not lead to the final decision to accept or reject a message. The degree to which central and peripher al cues affect the attitude of an individual depends on the level of motivation and/or message processing capability of the message recipient . If an individual is motivated and/or able to process a message, the central cues are more influential in

PAGE 61

. 50 changing his/her attitude, and in turn, affecting his/her judgment. On the other hand, if an individual is less motivated and/or able to process a message, the peripheral cues are more influential in changing his/her attitude, and thus leading him/her to accept or reject the message ( Angst and Agarwal 2009 ; Bhattacherjee and Sanford 2006 ; P etty and Cacioppo 1986 ) . ELM and Review Usefulness As a technological component of online review systems, the social voting feature, the main focus of this study, reviews. In particular, the social voting feature in Yelp can be seen as a technological component allowing users to express their views on the credibility of the online reviews themselves ( Baek et al. 2012 ) . Re view credibility, as the flip side of the perceived usefulness of online reviews, is among the most important concerns in online review research, revealing that reviews with higher credibility exert a greater influence in the adoption of eWOM (Baek et al. 2012; Cheung et al., 2009; Cheung et al., 2012). Prior studies examine perceived credibility of online reviews as a cognitive based attitude construct ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ; Fang 2014 ) peripheral cues to investigate the antecedents of the review credibility. Consistent with these prior studies, and by considering the notion that online consumer reviews contain both periphera l and central cues ( Mousavizadeh et al. 2015 ) , we believe ELM provides a suitable lens for investigating the drivers of review usefulness (as a notion of review credibility) in our study . Moreover, based on our literature review (refer to Table 1 in the Appendix section), prior studies investigate how the factors related to review text, as well as the

PAGE 62

. 51 attributes of reviewers, affect review usefulness. One of the major limitations of these studies is neglecting how the inconsistenc y between the central and peripheral cues of ELM affects the usefulness of online reviews. This notion is important from a theoretical perspective as ELM posits that persuasion can be simultaneously executed through the central and peripheral cues ( Angst and Agarwal 2009 ; Bhattacherjee and Sanford 2006 ; Petty and Cacioppo 1986 ) . To understand this phenomenon , in our research model, we theorize review consistency as a two factor measure, which reflects the consistency between the central cue (i.e., review text) and the peripheral cue (i.e., numerical star rating). we also conside r review peripheral cues) as theoretically important variables in our research model. Review Sentiment Polarity Past studies have examined the effect of review text as the central cue of ELM on review credibility ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Recent studies examining the content of online reviews find that the emotions embodied in the review text can substantially influence the way a review is processed ( Salehan and Kim 2014 ; Salehan and Kim 2016 ; Yin et al. 2014b ) . These findings are consistent with the idea that emotion s expressed in a review text perform as a source of social information ( Van Kleef 2010 ) . In addition, a recent study by Yin et al. (2014) shows that negative emotions such as anger and anxiety positively influence the usefulness of online reviews. Past studies attribute the positive impact of a negat ( Cao et al. 2011 ; Kuan et al. 2015 ) . Thi s negativity bias implies that negative information is perceived as more informative and diagnostic than positive information ( Baumeister et al.

PAGE 63

. 52 2001 ; Herr et al . 1991 ) . Additionally, negative reviews are more vivid than positive ones, and in an online environment where positive reviews are salient, negative reviews are more likely to gain the attention of consumers ( Herr et al. 1991 ; Kuan et al. 2015 ) . Moreover, negative information reflects a low quality of products, however, positive information can be attributed to both low and high quality products ( Skowronski and Carlston 1987 ) . Conversely, studies that find that positive ratings are more useful ( Carlson and Guha 2010 ; Chevalier and Mayzlin 2006 ) argue that, once consumers are p redisposed to a product prior to a purchase, they ignore negative ratings because such ratings do not comply with the Synthesizing the prior research, we believe that, in the context of a service industry, negativity bias is a bett er explanation for how a review with negative se n timent polarity affects different from product experiences. That means, service experiences tend to involve a high level of su bjectivity and a high volume of online reviews, leading consumers to find negative information more diagnostic and vivid, thus resulting in negative reviews receiving more useful votes. Therefore, we suggest following hypothesis: H 1 : Reviews with negative sentiment polarity are more likely to receive useful votes than positive ones. Review Rating The numerical star rating of an online review is a peripheral factor that reflects a ( Mudambi and Schu ff 2010 ) . A one or two star rating indicates a negative view of a product or service, a three star rating reflects a moderate view, and a four or five star rating reflects a positive view of a product or service

PAGE 64

. 53 ( Mudambi and Schuff 2010 ) . Consistent with neg ativity bias, prior studies report a positive effect of negative ratings on review usefulness ( Cao et al. 2011 ; Kuan et al. 2015 ; Sen and Lerman 2007 ; Willemsen et al. 2011 ) . In line wit h these prior studies, we suggest that: H 2 : The numerical star rating of online reviews has an inverse relationship with the review usefulness. Source Credibility Information available in profiles of online community users, especially those on online cons umer review sites, provides valuable information about the message source ( Baek et al. 2012 ; Forman et al. 2008 ; Kuan et al. 2015 ) of the credibility of a reviewer, is an i mportant ELM peripheral cue that has been studied in the past ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ; Kuan et al. 2015 ) . Source credibility has been shown to have a positive effect on the credibility of eWOM in online consumer review platforms ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Past studies have also found source credibility to have a positive effect on the usefulness of online reviews. Forman et al. (2008) found that revealing the identity of a reviewer has a positive effect on the usefulness of reviews. More recent studies ( Baek et al. 2012 ; Kuan et al. 2015 ) have also found that platform based signals of source credibility, such as the top revie wer badge on Amazon, have a positive effect on the usefulness of online reviews. As Yelp uses a reviewer badge is a good proxy of reviewer credibility. Thus, we po sit that: H 4 : Reviews from Elite badge members are more likely to receive useful votes than those without badges. Review Consistency Past studies have examined the effect consistency of online reviews on review usefulness. ( Baek et al. 2012 ; Cheung et a l. 2012 ; Cheung et al. 2009 ; Quaschning et al.

PAGE 65

. 54 2015 ; Zhang and Watts 2003 ) . These prior studies examine the extent to which a recommendation by a reviewer is consistent with those of other reviewers concerning the same produ ct or service evaluation. Baek et al. ( 2012 ) have evaluated the consistency of reviews in terms of the agreement rating for the associated product/service and have found that a conflict between an individual rating and the aggregated rating has a negative impact on the usefulness of an online review. While these prior studies have examined the consistency across reviews and in terms of agreement between two peripheral cues (i.e., a reviews rating and the average rating), less attention has been paid to the role of consistency between a review text and the corresp onding numerical rating (i.e., the central and peripheral cues of online reviews) and its effect on review usefulness. According to Tang and Guo (2015), eWOM communication begins when a consumer develops attitudes toward a product/service based on his/her consumption experience. The consumer then incorporates those attitudes into a form of review through both textual commentaries (i.e., a central cue) and an associated numerical nto cues contained in the eWOM text is a communication process called encoding or expression. On the other hand, decoding or impression process happens when consumers read online reviews and make the decision whether to accept ing or reject the review by ev aluating both verbal (i.e., review text) and non verbal (i.e., review rating) features of online reviews. Since both review service, reading the review text itself should lead the consumers to the same numerical rating during the decoding process ( Mudambi et al. 2014 ) . Therefore, the inconsistency between review text and the corresponding review rating is considered as a source of conflicting

PAGE 66

. 55 information for the reader, resulting in ambiv alent feelings for the reader ( Chang 2012 ; Chang 2013 ; Chang 2016 ; Monteith 1996 ) , triggering discomfort ( Nordgren et al. 2006 ) and consequently resulting in a lower perceived credibility of the review . Therefore, we suggest that: H 4 : The consistency between review text and its attendant star rating has a positive effect on review usefulness. Research Methodology Data Collection Data in this study was collected from yelp.com. Yelp has been widely recognized as one of the major online consumer review sites on service industries, especially restaurants. We used Data Toolbar, a commercial web scraping software, to collect the consumer reviews. Following Salehan and Kim (2014) , we selected restaurants that have received at least 10 0 reviews. In total, 2 , 634 usable reviews were collected from 20 restaurants in the metropolitan area of one of the major cities in the western United States in 2015. Measures The dependent variable in this study is review usefulness. It is measured as the number of useful votes a review received. The independent variables of this study are review sentiment polarity, review rating, source credibility and review consistency. In addition to them, control variables include the review length, review longevity, average restaurant rating, Measure of the Review Consistency Review consistency in this study is mea sured by using two different methods. As the first method, we employ the unsupervised classification method to examine the consistency

PAGE 67

. 5 6 between the polarity of the review sentiment and the number of stars attached to the review . As the second method, we emp loy the machine learning techniques to predict the consistency of the review. These methods are described in detail below. Method 1: Unsupervised Classification Generally, there are two main approaches to the problem of extracting sentiment automatically ( Pang and Lee 2008 ) . The first method, the lexicon based (unsupervised) approach, involves calculating the polarity of sentiment based on the words and phrases u sed in the text using natural language processing and computational linguistic approaches. This is in contrast to the second method, the supervised approach, wherein researchers train classifier algorithms based on predefined labels of the review sentiment ( Taboada et al. 2011 ) . Sentiment mining is usually done using the lexicon based approach because of its efficiency and scalability ( Bai 2011 ) . In this study, we used Semantria, which has also been used in past studies ( Charissiadis and Karacapilidis 2015 ; Kim et al. 2015 ; Lawrence 2014 ) . Semantria is a commercial software that provides an Excel plug in for conducting an automatic sentiment analysis on text data . 4 Semantria breaks each document i nto different Part of Speech (PoS) and then identifies the sentiment bearing phrases, which earn a logarithmic scale ranging between 1 and 1. Semantria combines the scores of those phrases to determine the overall sentiment polarity of the document, which can be labeled as either positive, negative, or neutral. In our study , we treated review consistency as the consistency between a revie expressed in the review text . We labeled reviews that received 1 or 2 star(s) as negative, 3 stars as neutral, and 4 or 5 stars as positive. The review consistency variable was treated as a binary var iable; that 4 https://semantria.com/support/resources/technology

PAGE 68

. 57 Method 2: Supervised Classification As the second method to measure the review consistency, a machine learning technique (a supe rvised approach) was used, which involves training of classifier algorithms. To perform this supervised classification, the Graphlab module in Python was used. Graphlab is an extensible machine learning framework that includes rich libraries for data trans formation and manipulation as well as a task oriented machine learning toolkit for creating, evaluating, and improving the machine learning models. Before training the classifiers, we parsed the text data to remove unwanted characters and punctuation. T o t rain by the term by frequency matrix, each column represents a unique word that appears across all reviews, and each row refers to each individual review text. Each cell in the matrix represents th e number of times that a term (column) appears in a particular review text (row). Since somewhat meaningless words eliminated from the matrix in order to reduce t he number of columns. While term by frequency matrix weighting has been used to train classifier algorithms in past studies ( Ngo Ye and Sinha 2014 ) , using term frequency alone cannot effectively distinguish among reviews ( Cao et al. 2011 ) because a term that appears commo nly in one type of review may also appear in other types. For instance, in Yelp this problem, the term frequencies were adjusted by TF IDF weighting (term frequency inverse

PAGE 69

. 58 IDF value increases comparably to the number of times a word appears in a review, but it is offset by the frequency of the word in all reviews ( Arazy and Woo 2007 ; Cao et al. 2011 ; Salton et al. 1975 ) . The TF IDF has been reported to increase information retrieval precision by up to 70% when compared to the term by frequency matrix ( Salton et al. 1975 ) . Equation 1 expresses the standard TF IDF weighting ( Arazy and Woo 2007 ; Cao et al. 2011 ; Salton et al. 1975 ) : In the above equation, is the weighted frequency of term in document ; is the frequency of term in document and , where is the number of the documents and is the frequency of term in the documents. In the next step, 400 records were randomly selected to train and test the performance of the classifiers. Two undergraduate students were asked to rea d the online reviews and examine the consistency between the review text and the associated numerical star rating. in the review text i s consistent with the numeri otherwise. The Cohen's kappa ( Carletta 1996 ) inter rater reliability measure was found to exceed 0.7, indicating a probability that the agreed understanding between the two student coders was significantly higher than what can be obtained by chance ( Krippendorff 2012 ; Landis and Koch 1977 ) . The coders next met to resolve conflicting ratings until the overall agreements conver ged. The review consistency labels provided through the coding process were used to train and test the performance of the classifiers in the machine learning step. In addition, we sought to evaluate the performance of support vector machines (SVMs) and ra ndom forest classifiers in predicting the review consistency labels. These

PAGE 70

. 59 classifiers have previously been used to study the usefulness of online reviews ( Ghose and Ipeirotis 2011 ) . To predict the review consistency labels, the 400 selected records were randomly split into 70% of training data and 30% of test data to test their performance. The TF IDF matrix and the numerical star rating of the reviews were used as the set of features to train the classifiers based on the training set. Further, accuracy metrics were used to evaluate the perform ance of the classifiers based on the test set. Our results show that the random forest algorithm (with 91% accuracy) outperforms the SVM classifier (with 61% accuracy), which is consistent with the results of the past study ( Ghose and Ipeirotis 2011 ) . Finally, the model built on random forest classifiers was used to predict the unknown review consistency labels in the remaining 2234 records. Those 2234 re cords were later used to empirically test the hypotheses using ordered logit and negative binomial regression. Figure 3 summarizes the steps involved in predicting the review consistency using the machine learning approach. Other Measures Source credibili ty, which refers to the credibility of the reviewer ( Baek et al. 2012 ; Kuan et al. 2015 ) , is operationalized as a binary va riable. If the reviewer was an Elite badge measured as the number of words contained in the review ( Baek et al. 2012 ; Salehan and Kim 2014 ; Salehan and Kim 2016 ) . Review longevity was measured by the number of months elapsed from the date when the review was posted. Rating incon sistency was ratings of the corresponding restaurant. The review sentiment polarity, ranging from 1 to 1, was used to understand whether a review was negative, neutra l or positive. In addition,

PAGE 71

. 60 number of friends and l variables in our study. Empirical Models In this study, we are interested in understanding (1) why a review receive s a useful vote, and (2) what factors contribute for reviews to receive useful votes ( Cao et al. 2011 ) . The ordered logit model, as an extension of the binary logit model, can accommodate a dependent variable that has more than two categories. In our study, the dependent variable is the number of useful votes a review has received from other online users . We followed the Cao et al. (2011) approach to categoriz ing the number of useful votes. Table 4 shows the reviews by the number of useful votes in the eight categories. Moreover, the dependent variable of our study is the n umber of useful votes received by other online users. As is shown in Table 4, a large proportion of the reviews (50.21%) did not receive a single useful vote, which shows the overdispersed variance of useful votes. Therefore, we used negative binomial regr ession, one of the Poisson model variations ( Greene 1994 ; Schindler and Bickart 2012 ) . To ensure that negative binomial regression was suitable for the purposes of this study over the Poisson model, we tested whether the overdispersion parameter was significantly different from zero. Our results gave a p value < 0.001, which confirms the existence of overdispersion in our dataset, validating that it was appropriate to use negative binomial regression over the Poisson model. Therefore, our model was estimated by using both an ordered logit and a negative bi nomial regression. We built 5 models, in order to check whether adding a new variable of review consistency significantly improves the variance of the review usefulness. The basic model (model 1 in

PAGE 72

. 61 Table 5) contains only control variables. Model 2 estimate s the full model excluding the review consistency variables. Model 3 estimates the full model including the unsupervised variable of the review consistency. Model 4 estimates the full model including the supervised variable of the review consistency, and m odel 5 estimates the full model including both the unsupervised and supervised variables of the review consistency. The below equations show our regression models. In all the equations, the variables in italics are log transformed. Model 1: Control variab les: Average Restaurant Rating Model 2: All the variables excluding review consistency variables: Average Restaurant Rating Model 3: All the variables inclu ding the unsupervised measure of the review consistency variable: Average Restaurant Rating

PAGE 73

. 62 Model 4: All the variables including the supervised measure of the review consistency variable: Average Restaurant Rating Model 5: All the variables including the unsupervised and supervised measures of the review consistency variable: Average Restaurant Rating

PAGE 74

. 63 Discussion Models 3 column in Table 4 is used to evaluate our hypotheses. The first hypothesis (H 1 ) examined the effect of review sentiment polarity on the review usefulness. We expected that a negative sentiment polarity would positively influence the review usefulness. However, we did not find support for this hypothesis. One possible explanation for this find ing might be related to the effect of argument quality of review usefulness ( Cheung et al. 2012 ; Cheung et al. 2009 ) . In other words, it is possible that a review with low numerical ratings (i.e., one or two star ratings) gains the attention of consumers and persuades them to arguments about why he/she assigns a low numerical rating to the restaurant. In other words, the reviewe r might be anxious or angry and tends to express his/her emotions by writing a review. However, his/her emotions weaken the rational arguments in the review text ( Wang 2006 ) , resulting in making his/her review less useful. Our second hypothesis (H 2 ) examined the effect of the rev usefulness. Our results show that the numerical star rating of a review has a negative effect on the usefulness of online reviews ( = 0.17, p<0.01), and its exponentiated odds ratio 5 is 5 To better interpret the results of ordered logit in our analysis we used the value of odds ratio. The odds ratio is better than The odds ratios can be obtained by exponentiating the ordered logit coefficients (i.e., )) ( Field 2009 ) . The odds ratio is defined as: . Given all other predictors held constant, a value greater than 1 indicates that as the predictor increases by one unit, the odds of outcome increases. On the other hand, a value less than one indicates that as the predictor increases by one unit, the od ds of outcome occurring decreases. In other words, the negative value of indicates a decrease in the likelihood of the outcome occurring given one unit increase in the predictor and a positive value of indicates an increase in the likelihood of the outcome occurring given one unit increase in the predictor. For instance, in H1, = 0.17 and the odds ratio is represented as , indicating 16% decrease in the likelihood of receiving a useful vote. On the other hand, the values of coefficient is positive in H2, H3 and H4, indicating an increase in the likelihood of receiving useful votes given one unit increase in the predictors.

PAGE 75

. 64 represented as . That means, given the other variables in the model held constant, for a one unit increase in review numerical rating, the likelihood of receiving useful votes decreases by 16%. Our finding confirms the existence of a negativity bias among online reviews ( Kuan et al. 2015 ) . As reported by ( Gao et al. 2015b ; Hu et al. 2009 ) , the majority of online revi ews are positive, which reflects the typical J shaped distribution of star rating. Consistently, in our dataset, most of the reviews (70%) were found to have either four or five star ratings (refer to Figure 4). Thus, since most of the reviews in our datas et are positive, negativity bias suggests that negative information is generally deemed more diagnostic than positive information ( Kuan et al. 2015 ) . The third hypothesis (H 3 ) contemplated the effect of source credibility on review usefulness. Our results show that a review by an Elite badge member is more likely to receive a usefulness vote ( =0.84, p<0.01). The odds ratio is .That means, given all other variables held constant, an elite badge member is 2.31 times more likely to receive a useful vote than a non elite badge member. Thus, H 3 was supported. Our results also confirm the findings of past studies ( Baek et al. 2012 ; Kuan et al. 2015 ) reporting the positive effect of platform generated measures of source credibility (i.e., top reviewer rank on Amazon) on review usefulness. Our fourth hypothesis (H 4 ) examined the effect of review consistency on the review usefulness. In model 3, we inc luded the unsupervised measure of review consistency and examined its effect on the review usefulness. The result of the ordered logit regression indicates a positive effect of the unsupervised measure of review consistency on review usefulness ( =0.78, p<0.01). The odds ratio is . That means, given all other variables held constant, consistent reviews are 2.18 times more likely to receive useful

PAGE 76

. 65 votes than inconsistent reviews. In model 4, we examined the effect of supervised measure o f review consistency on review usefulness, and our results support its positive effect on review usefulness ( =0.83, p<0.01). Therefore, our models show strong support for H 4. In model 5, we included both the unsupervised and supervised variables of re view consistency. Our result yields the existence of multicollinearity problem, because of the high correlation between the unsupervised and supervised variables of review consistency. To address the problem, we further examine the variance inflation facto r (VIF) of the independent variables based on model 5, and the result shows high VIF values for Unsupervised Review Consistency (VIF=5.2) and Supervised Review Consistency (VIF=8.3). To check whether adding the review consistency variable in model 3 would enhance the R 2 above and beyond model 1 and model 2, we run a hierarchical regression on the successive models. As presented in Table 5, our result shows that adding the review consistency variable to model 3 significantly improves the variance in the usef ulness of online reviews above and beyond model 1 and model 2. However, this result does not hold for model 4 and model 5 because of the high correlation between supervised and unsupervised measures of review consistency. Given the focal variable (i.e., re view consistency) that we developed and tested in this study, it confirms the validity of ELM in the online review context. In other words, online reviewers tend to process central cue (i.e., verbal review text) and peripheral cue (i.e., non verbal star ra ting) simultaneously, not in isolation. Post hoc Analysis Our post hoc analysis is theoretically motivated by two important reasons. First, prior studies on eWOM communication process have maintained that the credibility of a reviewer

PAGE 77

. 66 (e.g., if the reviewer is an elite badge member or no) is the most important peripheral cue that affects eWOM credibility. Likewise, Baek et al. (2012) and Kuan et al, (2015) find that the iewer in Amazon has a positive effect on review usefulness. Second, ELM states that individuals start to process information by incorporating peripheral cues, and then, depending on the quality or strength of peripheral cues , they might be motivated to process central cues ( Mousavizadeh et al. 2015 ; Petty and Cacioppo 2012 ; Sundar and Kim 2005 ) . Re flecting the prior studies and by considering that elite badge member turns out to be a strong peripheral cue as a proxy of the reviewer credibility (refer to Table 5), we believe examining the different structural pattern of review usefulness based on whe ther a reviewer is credible (i.e., elite badge =1) or not credible (i.e., elite badge =0) will shed more light on how consumers process central and peripheral cues. As reported in table 6, our analysis reveals that elite badge members and non elite badge members show different patterns in processing central and peripheral cues. To text (i.e., review length where =0.54 , 1.71), and disregard other peripher al cues, because the presence of elite badge itself works as a strong peripheral cue that warrants the credibility of the reviewers. In other words, the existence of elite badge member cue overrides the signals of other peripheral cues, and direct online u content of online reviews. Therefore, reviews that elaborate more on a service (i.e., lengthier reviews) are more likely to receive useful votes with an expectation of online users that reviews written by elite badge members would provide credible information nuggets about the service quality of reviewed restaurants ( Baek et al. 2012 ) . In contrast, if a review is

PAGE 78

. 67 written by a non elite badge member, online us ers would have struggled to process all other peripheral cues to assess if an online review would be credible. It is reflected by statistically significant effects of various peripheral cues (i.e., review rating, rating inconsistency, review longevity, rev review length), on the review usefulness. Interestingly, Table 6 shows that the review consistency variable has different effects on review usefulness. That means, in the c ase of a review written by an elite badge member, the effect of review consistency turns out to be insignificant. However, if a review is written by a non elite badge member, the review consistency becomes a significant driver of review usefulness ( =0.62 , ). For the different effects of review consistency, we effect of review consistency can be attributable to the sweeping effect of elite badge m ember as a strong peripheral cue. That means, as the existence of an elite badge itself guarantees the credibility of the review, online users begin to read the reviews even without bothering themselves to evaluate if the review shows consistency between r eview text and review rating. However, as a review posted by non elite badge member does not guarantee the credibility of the reviewer, online users would have exerted extra cognitive effort to evaluate if the review shows consistency between review text a nd review rating. Second, the different effects of review consistency between elite and non elite member can be related with the different length of review (i.e., average number of words per review). That means, as elite badge members tend to write longer reviews that non elite badge members 6 , the former 6 While, average length of reviews (i.e., number of word per review) by elite badge member is 162, t he same by non elite badge members is 89.

PAGE 79

. 68 reviews would have made it harder for online users to evaluate the consistency between review text and review rating. Theoretical and Practical Contributions Theoretical Contributions This study contributes to prior review usefulness literature in two ways. First, ELM states that central and peripheral cues of a message are processed jointly, not in isolation, by message recipients. Therefore, in the online review context, it makes more sense to consider review rating (i.e., the peripheral cue) and the review text (i.e., the central cue) simultaneously. However, prior studies ( Baek et al. 2012 ; Forman et al. 2008 ; Ku an et al. 2015 ; Mudambi et al. 2014 ; Salehan and Kim 2014 ; Yin et al. 2014a ) have focused on single factor measures (i.e., either review rating or review text) to explain their isolate d effect on review usef ulness. This study fills this theoretical gap by crafting and operationalizing a two factor variable (i.e., review consistency), which measures the consistency between review text and review rating (i.e., central and peripheral cues) and its effect on revi ew usefulness. Our analysis empirically demonstrates that our review consistency variable better explains the mechanism of online review usefulness above and beyond prior studies. Second, while prior research suggested that reviewer credibility is the mos t important peripheral cue of ELM, its mechanism has not been well explained ( Baek et al. 2012 ; Cheung et al. 2012 ; Cheung et al. 2009 ) . However, consistent with ELM, our post hoc analysis empirically demonstrates the mechanism that peripheral cue (i.e., elite badge) serves as signals that motivate online users to proc ess central cue (i.e., review text) in certain ways ( Kua n et al. 2015 ; Mousavizadeh et al. 2015 ; Petty et al. 2002 ) . Our results suggest that, in presence of reviewer credibility signal (i.e., an elite bade member in this study), online users

PAGE 80

. 69 move their focus to review text itself and other peripheral cues become less important. However, in the absence of the credible signal of the elite badge, online users exert extra cognitive efforts to mobilize other central and peripheral cues as a means to evaluate the information quality of online reviews. Practical Contributions Online retailers c an benefit from the result of this study by incorporating review consistency variable in sorting online reviews. Specifically, the unsupervised measure of review consistency can be integrated into the sorting algorithms of online review systems, because of its low cost implementation. An alternative approach is to examine the consistency between review text and review rating before a reviewer post a review. That means, the review system can be complemented with a real time alert component that can inform on line reviewers with the possible inconsistency between review text and review rating before he/she concludes the review. Limitations and Future Research on Yelp. Therefore, findings in this s tudy should be generalized with caution. Previous studies have shown the different effects of platforms ( Lee and Youn 2009 ) and product types ( Baek et al. 2012 ; Mudambi and Schuff 2010 ) on eWOM. Thus, more research needs to be done to know how different product types and platforms would affect the review usefulness vote. This study also opens a new avenue for how the text mining techniques can be used to examine the usefulness of online reviews. For instance, the argument h as been named as one the most important factors that influence the perceived credibility of online reviews ( Cheung et al. 2012 ; Cheung et al. 2009 ) . Prior studies have measured argument quality as a latent

PAGE 81

. 70 variable by using data collected through surveys or exper iments ( A wad and Ragowsky 2008 ; Cheung et al. 2012 ; Cheung et al. 2009 ; Chu and Kamal 2008 ; Park et al. 2007 ) . These prior measure argument quality as a constr uct that reflects the degree to which an online review is believable, strong, persuasive, good and informative ( Cheung et al. 2012 ; Cheung et al. 2009 ) . One limitation of this view of argument quality is it is difficult in practice for online review providers to quantify these dimensions so that they can improve their recommendation systems. We believe, an alternative approach to quantify ing argument qua lity is to draw on the studies that suggest more quantifiable dimensions of argument quality . For instance, the recommendation consistency ( Baek et al. 2012 ) can be examined as the degree of consistency between the textual content of an online review and other reviews concerning the same product or service. In addition, the review accuracy ( Otterbacher 2009 ) can be studied as the degree of consistency between a review and the description of the corresponding product on online consumer review sites. Therefore, it would be fruitful for future studies to use text mining and natural language processing to examine the effect of these variabl es on the review usefulness. Concluding Remarks Quality online reviews play an important role in helping online consumers make an informed decision about their purchase. Thus, understanding the determinant factors of useful online reviews is important for both scholars and online consumer review providers. By using ELM as the theoretical lens and using machine learning and sentiment mining in combination with econometrics analysis, we crafted and operationalized a two dimens ional variable (i.e., review consistency), which measures the consistency between review text and review rating, and its effect on review usefulness. In theory, this study advances our

PAGE 82

. 71 understanding of ELM in online review context by providing empirical ev idence that the two factor based review consistency variable outperforms the single factor variables of previous studies in explaining how online reviews are considered as useful by online users. In addition, our results shed light on how consumers process central and peripheral cues in different ways to evaluate online reviews based on whether or not an online review is written by an elite badge member.

PAGE 83

. 72 FIGURES AND TABLES Figure III 1 . An example of a Rev iew on Yelp Figure III 2 . Research Model

PAGE 84

. 73 Figure III 3 . Supervised Measure of the Review Consistency Figure III 4 . The Distribution of Rating Based on Number of Reviews 114 228 323 724 845 0 100 200 300 400 500 600 700 800 900 1 2 3 4 5 Number of reviews Numerical Star Ratings

PAGE 85

. 74 Table III 1 . Descriptive Statistics Variables Description N Mean St.Dev Median Min Max USEFULNESS Number of useful votes a review receives 2,234 0.98 1.6 0 0 20 Elite Badge member Whether a reviewer is an Elite badge member 2,234 0 0.27 0 0 1 Unsupervised Review Consistency Unsupervised measure of the recommendation consistency 2,234 0.59 0.49 1 0 1 Supervised Review Consistency Supervised measure of the recommendation consistency 2,234 0.52 0.5 1 0 1 Number of reviews a restaurant has received 2,234 523.63 245.20 531 155 942 Average Restaurant Rating Average numerical star rating of a restaurant 2,234 3.83 0.38 4 3 4 Yelp 2,234 63.12 119.60 15 0 1,12 on Yelp 2,234 2.54 7.61 0 0 84 Number of reviews posted by a reviewer on Yelp 2,234 131.96 238.70 32.5 1 2,83 Number of tips posted by a reviewer on Yelp 2,234 22.33 76.23 1 0 1,32 Review Longevity Months elapsed after a review has been posted 2,234 25.33 20.27 19 0 111 Review Length Length of review text 2,234 108.69 98.17 80 2 953 Review Rating Numerical star ratings of a review 2,234 3.86 1.17 4.00 1 5 Rating Inconsistency The absolute difference between numerical star rating of review and the average star rating of the corresponding restaurant 2,234 0.84 0.71 1 0 3 Review Sentiment Polarity Review sentiment polarity 2,234 0.49 0.68 1 1 1

PAGE 86

. 75 Table III 2 . Correlation Matrix 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 USEFULNESS 1.00 2 Elite Badge Member 0.39 1.00 3 Unsupervised Review Consistency 0.09 0.01 1.00 4 Review Length 0.41 0.33 0.02 1.00 5 Average Restaurant Rating 0.04 0.14 0.15 0.17 1.00 6 0.03 0.13 0.01 0.12 0.42 1.00 7 Review Longevity 0.22 0.05 0.06 0.08 0.36 0.26 1.00 8 0.38 0.56 0.00 0.28 0.08 0.07 0.18 1.00 9 Friends 0.42 0.52 0.03 0.30 0.09 0.09 0.16 0.65 1.00 10 0.27 0.29 0.03 0.15 0.02 0.03 0.13 0.44 0.44 1.00 11 0.26 0.37 0.02 0.21 0.01 0.06 0.13 0.59 0.51 0.22 1.00 12 Supervised Review Consistency 0.06 0.04 0.87 0.03 0.23 0.08 0.11 0.00 0.05 0.04 0.01 1.00 13 Review Rating 0.05 0.07 0.33 0.07 0.32 0.09 0.10 0.03 0.06 0.03 0.01 0.51 1.00 14 Rating Inconsistency 0.03 0.16 0.09 0.01 0.12 0.02 0.05 0.14 0.13 0.07 0.09 0.18 .0.38 1.00 15 Review Sentiment Polarity 0.03 0.07 0.44 0.00 0.15 0.03 0.05 0.04 0.07 0.05 0.02 0.64 0.41 0.25 1.00 Table III 3 . Categories of Useful Vote s Number of useful votes Number of reviews Percentage (%) 0 1124 50.21 1 916 27.70 2 244 10.92 3 116 5.19 4 61 2.73 5 25 1.11 6 19 0.85 7 or more 26 1.16

PAGE 87

. 76 Table III 4 . Econometric Analyses, Ordered Logit Variables Model 1 Model 2 Model 3 Model 4 Model 5 Elite Badge Member 0.85*** 0.84*** 0.82*** 0.83*** Review Rating 0.10** 0.17*** 0.22*** 0.19*** Review Sentiment Polarity 0.29** 0.1 0.04 0.03 Unsupervised Review Consistency 0.78*** 0.53*** Supervised Review Consistency 0.83*** 0.3 Review Length 0.53*** 0.50*** 0.50*** 0.51*** 0.51*** Average Restaurant Rating 0.1 0.1 0.07 0.05 0.06 Number of Reviews 0.98*** 0.99*** 0.94*** 1.02*** 0.97*** Review Longevity 022*** 0.26*** 0.27*** 0.27*** 0.27*** Rating Inconsistency 0.39** 0.40*** 0.37*** 0.37*** 0.37*** Number of Reviews 0.12** 0.03 0.05 0.05 0.05 Number of Friends 0.15** 0.13*** 0.12*** 0.12*** 0.12*** Number of Tips 0.05** 0.04* 0.04* 0.04* 0.04* Number of Followers 0.04 0.01 0.02 0.02 0.02 Observation 2,234 2,234 2,234 2,234 2,234 R 2 0.3 0.32 0.35 0.34 0.35 Chi2 744.85*** 805.10*** 865.87*** 860.04*** 868.28*** Note: *p<0.1; **p<0.05; ***p<0.01

PAGE 88

. 77 Table III 5 . Comparing Successive Models RSS DF Sum of Square F statistics P value Model 1 : Control Variables 3015.7 Model 2 : Model 1+Independent Variables (Excluding Review Consistency Variables) 2957.5 3 58.19 14.81 < 0.001 Model3 : Model 2+ Unsupervised Review Consistency 2910.8 1 46.69 35.65 < 0.001 Model 4 : Model 2+ Supervised Review Consistency 2908.3 0 2.538 Model 5 : Model 2+Unsupervised Review Consistency + Supervised Review Consistency 2905.9 1 2.418 1.8464 0.1743 Note: *p<0.05; **p<0.01; ***p<0.001 Table III 6 . Post hoc Analysis Variables Elite Badge Members Non Elite Badge Members Logit Logit R eview Rating 0.22 0.14** Review Sentiment Polarity 0.32 0.04 Unsupervised Review Consistency 20.65 0.62*** Review Length 0.54*** 0.38*** Average Restaurant Rating 0.13 0.08 1.93** 0.68*** Review Longevity 0.13 0.31*** Rating Inconsistency 0.31 0.29*** 0.13 0.004 0.17 0.10*** 0.06 0.01 0.03 0.13* Observation 596 1,638 Note: *p<0.1; **p<0.05; ***p<0.01

PAGE 89

. 78 REFERENCES Adams, D. A., Nelson, R. R., and Todd, P. A. 1992. "Perceived Usefulness, Ease of Use, and Usage of Information Technology: A Replication," MIS quarterly (16:2 ) , pp. 227 247. Angst, C. M., and Agarwal, R. 2009. "Adoption of Electronic Health Records in the Presence of Privacy Concerns: The Elaboration Likelihood Model and Individual Persuasion," MIS quarterly (33:2), pp. 339 370. Arazy, O., and Woo, C. 2007. "Enhancing Informat ion Retrieval through Statistical Natural Language Processign: A Study of Collocation Indexing," MIS quarterly (24:2 ) , pp. 525 546. Arnold, H. J., and Feldman, D. C. 1981. "Social Desirability Response Bias in Self Report Choice Situations," Academy of Man agement Journal (24:2), pp. 377 385. Awad, N. F., and Ragowsky, A. 2008. "Establishing Trust in Electronic Commerce through Online Word of Mouth: An Examination across Genders," Journal of Management Information Systems (24:4), pp. 101 121. Baek, H., Ahn, J., and Choi, Y. 2012. "Helpfulness of Online Consumer Reviews: Readers' Objectives and Review Cues," International Journal of Electronic Commerce (17:2), pp. 99 126. Bai, X. 2011. "Predicting Consumer Sentiments from Online Text," Decision Support Systems (50:4), pp. 732 742. Bakshy, E., Eckles, D., Yan, R., and Rosenn, I. 2012. "Social Influence in Social Advertising: Evidence from Field Experiments," Proceedings of the 13th ACM Conference on Electronic Commerce : ACM, pp. 146 161. Barclay, D., Higgins, C. , and Thompson, R. 1995. "The Partial Least Squares (Pls) Approach to Causal Modeling: Personal Computer Adoption and Use as an Illustration," Technology studies (2:2), pp. 285 309. Baumeister, R. F., Bratslavsky, E., Finkenauer, C., and Vohs, K. D. 2001. "Bad Is Stronger Than Good," Review of general psychology (5:4), p. 323. Berger, J., and Milkman, K. L. 2012. "What Makes Online Content Viral?," Journal of marketing research (49:2), pp. 192 205. Bhattacherjee, A., and Sanford, C. 2006. "Influence Process es for Information Technology Acceptance: An Elaboration Likelihood Model," MIS quarterly (30:4), pp. 805 825. Brown, J. J., and Reingen, P. H. 1987. "Social Ties and Word of Mouth Referral Behavior," Journal of Consumer research ), pp. 350 362.

PAGE 90

. 79 Cao, Q., Du an, W., and Gan, Q. 2011. "Exploring Determinants of Voting for the Decision Support Systems (50:2), pp. 511 521. Carletta, J. 1996. "Assessing Agreement on Classification Tasks: The Kappa Stat istic," Computational linguistics (22:2), pp. 249 254. Carlson, K., and Guha, A. 2010. "The Ratings Paradox: Why We Prefer Reading Negative Reviews, but Then Subsequently Rate These Reviews as Less Useful," Advances in Consumer Research (37). Casteleyn, J., Mottart, A., and Rutten, K. 2009. "How to Use Data from Facebook in Your Market Research," International Journal of Market Research (51:4), pp. 439 447. Chang, C. 2012. "Ambivalent Attitudes in a Communication Process: An Integrated Model," Human Communication Research (38:3), pp. 332 359. Chang, C. 2013. "The Influence of Ambivalence toward a Communication Source: Media Context Priming and Persuasion Polarization," Communication Research ), p. 0093650213495408. Chang, C. 2016. "Responses to C onflicting Information in Computer Mediated Communication: Gender Difference as an Example," new media & society (18:1), pp. 5 24. Charissiadis, A., and Karacapilidis, N. 2015. "Strengthening the Rationale of Recommendations through a Hybrid Explanations B uilding Framework," in Intelligent Decision Technologies . Springer, pp. 311 323. Cheung, C. M. Y., Sia, C. L., and Kuan, K. K. 2012. "Is This Review Believable? A Study of Factors Affecting the Credibility of Online Consumer Reviews from an Elm Perspective," Journal of the Association for Information Systems (13:8), p. 618. Cheung, C. M., Lee, M. K., and Rabjohn, N. 2008. "The Impact of Electronic Word of Mouth: The Adoption of Online Opinions in Online Customer Communities," Internet Research (18:3), pp. 229 247. Cheung, M. Y., Luo, C., Sia, C. L., and Chen, H. 2009. "Credibility of El ectronic Word of Mouth: Informational and Normative Determinants of on Line Consumer Recommendations," International Journal of Electronic Commerce (13:4), pp. 9 38. Chevalier, J. A., and Mayzlin, D. 2006. "The Effect of Word of Mouth on Sales: Online Book Reviews," Journal of marketing research (43:3), pp. 345 354. Chin, W. W. 1998a. "Commentary: Issues and Opinion on Structural Equation Modeling." JSTOR.

PAGE 91

. 80 Chin, W. W. 1998b. "The Partial Least Squares Approach to Structural Equation Modeling," Modern method s for business research (295:2), pp. 295 336. Chin, W. W., Marcolin, B. L., and Newsted, P. R. 2003. "A Partial Least Squares Latent Variable Modeling Approach for Measuring Interaction Effects: Results from a Monte Carlo Simulation Study and an Electronic Mail Emotion/Adoption Study," Information systems research (14:2), pp. 189 217. Chu, S. C., and Choi, S. M. 2011. "Electronic Word of Mouth in Social Networking Sites: A Cross Cultural Study of the United States and China," Journal of Global Marketing (24 :3), pp. 263 281. Chu, S. C., and Kamal, S. 2008. "The Effect of Perceived Blogger Credibility and Argument Quality on Message Elaboration and Brand Attitudes: An Exploratory Study," Journal of Interactive Advertising (8:2), pp. 26 37. Chu, S. C., and Kim, Y. 2011. "Determinants of Consumer Engagement in Electronic Word of Mouth (Ewom) in Social Networking Sites," International journal of Advertising (30:1), pp. 47 75. Connors, L., Mudambi, S. M., and Schuff, D. 2011. "Is It the Review or the Reviewer? A Mu lti Method Approach to Determine the Antecedents of Online Review Helpfulness," System Sciences (HICSS), 2011 44th Hawaii International Conference on : IEEE, pp. 1 10. Cooper, S. 2015. "Facebook Is Leading the Way in Social Commerce." Corson, Y., and Verrie r, N. 2007. "Emotions and False Memories Valence or Arousal?," Psychological Science (18:3), pp. 208 211. Davis, F. D. 1989. "Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology," MIS quarterly ), pp. 319 340. Ebermann, J., Stanoevska Slabeva, K., and Wozniak, T. 2011. "Influential Factors of Recommendation Behaviour in Social Network Sites an Empirical Analysis," ECIS . Edell, J. A., and Burke, M. C. 1984. "The Moderating Effect of Attitude toward an Ad on Ad Ef fectiveness under Different Processing Conditions," NA Advances in Consumer Research Volume 11 ). eMarketer. 2010. "The Role of Customer Product Reviews,"). eMarketer.com. 2014. "Facebook Is No. 1 for Social Commerce." Eroglu, S. A., Machleit, K. A., and Da vis, L. M. 2003. "Empirical Testing of a Model of Online Store Atmospherics and Shopper Responses," Psychology & Marketing (20:2), pp. 139 150.

PAGE 92

. 81 Fang, Y. H. 2014. "Beyond the Credibility of Electronic Word of Mouth: Exploring Ewom Adoption on Social Network ing Sites from Affective and Curiosity Perspectives," International Journal of Electronic Commerce (18:3), pp. 67 102. Field, A. 2009. Discovering Statistics Using Spss . Sage publications. Folkes, V. S. 1988. "The Availability Heuristic and Perceived Risk, " Journal of Consumer research (15:1), pp. 13 23. Forgas, J. P. 1995. "Mood and Judgment: The Affect Infusion Model (Aim)," Psychological bulletin (117:1), p. 39. Forgas, J. P., and George, J. M. 2001. "Affective Influences on Judgments and Behavior in Org anizations: An Information Processing Perspective," Organizational behavior and human decision processes (86:1), pp. 3 34. Forman, C., Ghose, A., and Wiesenfeld, B. 2008. "Examining the Relationship between Reviews and Sales: The Role of Reviewer Identity Disclosure in Electronic Markets," Information Systems Research (19:3), pp. 291 313. Fornell, C., and Larcker, D. F. 1981. "Evaluating Structural Equation Models with Unobservable Variables and Measurement Error," Journal of marketing research ), pp. 39 50. Gao, G. G., Greenwood, B. N., Agarwal, R., and Jeffrey, S. 2015a. "Vocal Minority and Silent Majority: How Do Online Ratings Reflect Population Perceptions of Quality?," MIS quarterly (39:3), pp. 565 589. Gao, G. G., Greenwood, B. N., Agarwal, R., and McCullough, J. S. 2015b. "Vocal Minority and Silent Majority: How Do Online Ratings Reflect Population Perceptions of Quality?,"). Gefen, D., Benbasat, I., and Pavlou, P. 2008. "A Research Agenda for Trust in Online Environments," Journal of Management Inf ormation Systems (24:4), pp. 275 286. Ghose, A., and Ipeirotis, P. G. 2011. "Estimating the Helpfulness and Economic Impact of Product Reviews: Mining Text and Reviewer Characteristics," Knowledge and Data Engineering, IEEE Transactions on (23:10), pp. 149 8 1512. Granovetter, M. S. 1973. "The Strength of Weak Ties," American journal of sociology ), pp. 1360 1380. Greene, W. H. 1994. "Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial Regression Models," .Leonard N. Stern School of Business, New York Universit ).

PAGE 93

. 82 Gupta, P., and Harris, J. 2010. "How E Wom Recommendations Influence Product Consideration and Quality of Choice: A Motivation to Process Information Perspective," Journal of Business Research (63:9), pp. 1041 1049. Hanse n, S. S., and Lee, J. K. 2013. "What Drives Consumers to Pass Along Marketer Generated Ewom in Social Network Games? Social and Game Factors in Play," Journal of theoretical and applied electronic commerce research (8:1), pp. 53 68. Themselves on the Internet?," Journal of interactive marketing (18:1), pp. 38 52. Herr, P. M., Kardes, F. R., a nd Kim, J. 1991. "Effects of Word of Mouth and Product Attribute Information on Persuasion: An Accessibility Diagnosticity Perspective," Journal of consumer research (17:4), pp. 454 462. Hu, N., Zhang, J., and Pavlou, P. A. 2009. "Overcoming the J Shaped Distribution of Product Reviews," Communications of the ACM (52:10), pp. 144 147. Huang, A. H., Chen, K., Yen, D. C., and Tran, T. P. 2015. "A Study of Factors That Contribute to Online Re view Helpfulness," Computers in Human Behavior (48), pp. 17 27. Huang, J., Su, S., Zhou, L., and Liu, X. 2013. "Attitude toward the Viral Ad: Expanding Traditional Advertising Models to Interactive Advertising," Journal of Interactive Marketing (27:1), pp. 36 46. Jansen, B. J., Zhang, M., Sobel, K., and Chowdury, A. 2009. "Twitter Power: Tweets as Electronic Word of Mouth," Journal of the American society for information science and technology (60:11), pp. 2169 2188. Kaplan, A. M., and Haenlein, M. 2010. "U sers of the World, Unite! The Challenges and Opportunities of Social Media," Business horizons (53:1), pp. 59 68. Keenan, A., and Shiri, A. 2009. "Sociability and Social Interaction on Social Networking Websites," Library Review (58:6), pp. 438 450. Kim, J ., Naylor, G., Sivadas, E., and Sugumaran, V. 2015. "The Unrealized Value of Incentivized Ewom Recommendations," Marketing Letters ), pp. 1 11. Krippendorff, K. 2012. Content Analysis: An Introduction to Its Methodology . Sage. Kuan, K. K., Hui, K. L., Prasa rnphanich, P., and Lai, H. Y. 2015. "What Makes a Review Voted? An Empirical Investigation of Review Voting in Online Review Systems," Journal of the Association for Information Systems (16:1), p. 48. Landis, J. R., and Koch, G. G. 1977. "The Measurement o f Observer Agreement for Categorical Data," biometrics ), pp. 159 174.

PAGE 94

. 83 Lawrence, L. 2014. "Reliability of Sentiment Mining Tools: A Comparison of Semantria and Social Mention,"). Lee, J., and Lee, J. N. 2009. "Understanding the Product Information Inference Process in Electronic Word of Mouth: An Objectivity Subjectivity Dichotomy Perspective," Information & Management (46:5), pp. 302 311. Lee, J., Park, D. H., and Han, I. 2008. "The Effect of Negative Online Consumer Reviews on Product Attitude: An Informat ion Processing View," Electronic commerce research and applications (7:3), pp. 341 352. Lee, M., and Youn, S. 2009. "Electronic Word of Mouth (Ewom) How Ewom Platforms Influence Consumer Product Judgement," International Journal of Advertising (28:3), pp. 473 499. Intention to Go to the Festival: Applying an Extension of the Technology Acceptance Model," International Journal of Hospitality Management (31:3), pp. 819 827. Lenhart, A. 2009. "Adults and Social Network Websites,"). LePage, E. 2015. "A Long List of Social Media Advertising Stats," in: hootsuite . Leung, X. Y., Bai, B., and Stahura, K. A. 2015. "The Marketing Effectiveness of Social Media in the Hotel Indust ry a Comparison of Facebook and Twitter," Journal of Hospitality & Tourism Research (39:2), pp. 147 169. Liang, H., Saraf, N., Hu, Q., and Xue, Y. 2007. "Assimilation of Enterprise Systems: The Effect of Institutional Pressures and the Mediating Role of Top Management," MIS quarterly ), pp. 59 87. Lin, R., and Utz, S. 2015. "The Emotional Responses of Browsing Facebook: Happiness, Envy, and the Role of Tie Strength," Computers in human behavior (52), pp. 29 38. Liu, H. 2007. "Social Network Profiles as Taste Performances," Mediated Communication (13:1), pp. 252 275. Luarn, P., Yang, J. C., and Chiu, Y. P. 2015. "Why People Check in to Social Network Sites," International Journal of Electronic Commerce (19:4), pp. 21 46. MacKenzie, S. B., Lutz, R. J., and Belch, G. E. 1986. "The Role of Attitude toward the Ad as a Medi ator of Advertising Effectiveness: A Test of Competing Explanations," Journal of marketing research ), pp. 130 143. Mangold, W. G., and Faulds, D. J. 2009. "Social Media: The New Hybrid Element of the Promotion Mix," Business horizons (52:4), pp. 357 365.

PAGE 95

. 84 M cCroskey, J. C., Hamilton, P. R., and Weiner, A. N. 1974. "The Effect of Interaction Behavior on Source Credibility, Homophily, and Interpersonal Attraction," Human Communication Research (1:1), pp. 42 52. Monteith, M. J. 1996. "Contemporary Forms of Preju dice Related Conflict: In Search of a Nutshell," Personality and Social Psychology Bulletin (22:5), pp. 461 473. Mooney, C. Z., Duval, R. D., and Duval, R. 1993. Bootstrapping: A Nonparametric Approach to Statistical Inference . Sage. Mousavizadeh, M., Kooh ikamali, M., and Salehan, M. 2015. "The Effect of Central and Peripheral Cues on Online Review Helpfulness: A Comparison between Functional and Expressive Products,"). Mudambi, S. M., and Schuff, D. 2010. "What Makes a Helpful Review? A Study of Customer R eviews on Amazon. Com," MIS quarterly (34:1), pp. 185 200. Mudambi, S. M., Schuff, D., and Zhang, Z. 2014. "Why Aren't the Stars Aligned? An Analysis of Online Review Content and Star Ratings," 2014 47th Hawaii International Conference on System Sciences : IEEE, pp. 3139 3147. Ngo Ye, T. L., and Sinha, A. P. 2014. "The Influence of Reviewer Engagement Characteristics on Online Review Helpfulness: A Text Regression Model," Decision Support Systems (61), pp. 47 58. Importance," Business Wire ). Nordgren, L. F., Van Harreveld, F., and Van Der Pligt, J. 2006. "Ambivalence, Discomfort, and Motivated Information Processing," Journal of experimental social psychology (42:2), pp. 252 258. Mouth Effects on Online Survey Procedures," Communication (11:4), pp. 1085 1103. Okazaki, S. 2009. "Social Influence Model and Electroni c Word of Mouth: Pc Versus Mobile Internet," International Journal of Advertising (28:3), pp. 439 472. Otterbacher, J. 2009. "'Helpfulness' in Online Communities: A Measure of Message Quality," Proceedings of the SIGCHI Conference on Human Factors in Compu ting Systems : ACM, pp. 955 964. Pan, L. Y., and Chiou, J. S. 2011. "How Much Can You Trust Online Information? Cues for Perceived Trustworthiness of Consumer Generated Online Information," Journal of Interactive Marketing (25:2), pp. 67 74.

PAGE 96

. 85 Pang, B., and L ee, L. 2008. "Opinion Mining and Sentiment Analysis," Foundations and trends in information retrieval (2:1 2), pp. 1 135. Park, D. H., Lee, J., and Han, I. 2007. "The Effect of on Line Consumer Reviews on Consumer Purchasing Intention: The Moderating Role of Involvement," International Journal of Electronic Commerce (11:4), pp. 125 148. Pavlou, P. A. 2003. "Consumer Acceptance of Electronic Commerce: Integrating Trust and Risk with the Technology Acceptance Model," International journal of electronic commer ce (7:3), pp. 101 134. Pavlou, P. A., and Dimoka, A. 2006. "The Nature and Role of Feedback Text Comments in Online Marketplaces: Implications for Trust Building, Price Premiums, and Seller Differentiation," Information Systems Research (17:4), pp. 392 414. Pavlou, P. A., and Gefen, D. 2004. "Building Effective Online Marketplaces with Institution Based Trust," Information systems research (15:1), pp. 37 59. Petty, R., and Cacioppo, J. T. 2012. Communication and Persuasion: Central and P eripheral Routes to Attitude Change . Springer Science & Business Media. Petty, R. E., and Cacioppo, J. T. 1986. The Elaboration Likelihood Model of Persuasion . Springer. Petty, R. E., Priester, J. R., and Brinol, P. 2002. "Mass Media Attitude Change: Impli cations of the Elaboration Likelihood Model of Persuasion," Media effects: Advances in theory and research (2), pp. 155 198.2002. Qiu, L., Pang, J., and Lim, K. H. 2012. "Effects of Conflicting Aggregated Rating on Ewom Review Credibility and Diagnosticity: The Moderating Role of Review Valence," Decision Support Systems (54:1), pp. 631 643. Quaschning, S., Pandelaere, M., and Vermei r, I. 2015. "When Consistency Matters: The Effect of Valence Consistency on Review Helpfulness," Communication (20:2), pp. 136 152. Explo ratory Investigation across Three Services Categories," Electronic Commerce Research and Applications (11:6), pp. 548 559. Raguseo, E., and Vitari, C. 2017. "The Effect of Brand on the Impact of E Financial Performance," International Journa l of Electronic Commerce (21:2), pp. 249 269. Riegner, C. 2007. "Word of Mouth on the Web: The Impact of Web 2.0 on Consumer Purchase Decisions," Journal of advertising research (47:4), pp. 436 447.

PAGE 97

. 86 Riordan, M. A., and Kreuz, R. J. 2010. "Emotion Encoding and Interpretation in Computer Mediated Communication: Reasons for Use," Computers in human behavior (26:6), pp. 1667 1673. Salehan, M., and Kim, D. 2014. "Predicting the Performance of Online Consumer Reviews: A Sentiment Mining Approach." Salehan, M., an d Kim, D. J. 2016. "Predicting the Performance of Online Consumer Reviews: A Sentiment Mining Approach to Big Data Analytics," Decision Support Systems (81), pp. 30 40. Salehan, M., Kim, D. J., and Lee, J. N. 2015. "Antecedents, Processes and Consequences of Web Assurance Seals: A Meta Analysis Approach,"). Salton, G., Wong, A., and Yang, C. S. 1975. "A Vector Space Model for Automatic Indexing," Communications of the ACM (18:11), pp. 613 620. Schindler, R. M., and Bickart, B. 2012. "Perceived Helpfulness o f Online Consumer Reviews: The Role of Message Content and Style," Journal of Consumer Behaviour (11:3), pp. 234 243. Sen, S., and Lerman, D. 2007. "Why Are You Telling Me This? An Examination into Negative Consumer Reviews on the Web," Journal of interact ive marketing (21:4), pp. 76 94. Shih, H. p., Lai, K. h., and Cheng, T. 2013. "Informational and Relational Influences on Electronic Word of Mouth: An Empirical Study of an Online Consumer Discussion Forum," International Journal of Electronic Commerce (17 :4), pp. 137 166. Shimp, T. A. 1981. "Attitude toward the Ad as a Mediator of Consumer Brand Choice," Journal of advertising (10:2), pp. 9 48. Skowronski, J. J., and Carlston, D. E. 1987. "Social Judgment and Social Memory: The Role of Cue Diagnosticity in Negativity, Positivity, and Extremity Biases," Journal of personality and social psychology (52:4), p. 689. Söderlund, M., and Rosengren, S. 2007. "Receiving Word of Mouth from the Service Customer: An Emotion Based Effectiveness Assessment," Journal of r etailing and consumer services (14:2), pp. 123 136. User Interface Engineering ( http://www . uie. com/articles/magicbehindamazon/) ). Storbeck, J., and Clore, G. L. 2008. "Affective Arousal as Information: How Affective Arousal Influences Judgments, Learning, and Memory," Social and Personality Psychology Compass (2:5), pp. 1824 1843.

PAGE 98

. 87 Sun, T., Youn, S., Wu, G., and Kuntaraporn, M. 2006. An Exploration of Its Antecedents and Consequences," Communication (11:4), pp. 1104 1127. Sundar, S. S., and Kim, J. 2005. "Interactivity and Persuasion: Influencing Attitudes with Information and Involvement," Journal of interactive advertising (5:2), pp. 5 18. Sussman, S. W., and Siegal, W. S. 2003. "Informational Influence in Organizations: An Integrated Approach to Knowledge Adoption," Information systems research (14:1), pp. 47 65. Svensson , A. 2011. "Facebook the Social Newspaper That Never Sleeps a Study of Facebook Taboada, M., Brooke, J., Tofiloski, M., Voll, K., and Stede, M. 2011. "Lexicon Based Methods for Sentiment Analysis," Computational l inguistics (37:2), pp. 267 307. Tang, C., and Guo, L. 2015. "Digging for Gold with a Simple Tool: Validating Text Mining in Studying Electronic Word of Mouth (Ewom) Communication," Marketing Letters (26:1), pp. 67 80. Van Kleef, G. A. 2010. "The Emerging V iew of Emotion as Social Information," Social and Personality Psychology Compass (4:5), pp. 331 343. Vogt, J., De Houwer, J., Koster, E. H., Van Damme, S., and Crombez, G. 2008. "Allocation of Spatial Attention to Emotional Stimuli Depends Upon Arousal and Not Valence," Emotion (8:6), p. 880. Wang, X. T. 2006. "Emotions within Reason: Resolving Conflicts in Risk Preference," Cognition and emotion (20:8), pp. 1132 1152. Wang, Z. 2010. "Anonymity, Social Image, and the Competition for Volunteers: A Case Study of the Online Market for Reviews," The BE Journal of Economic Analysis & Policy (10:1). Wathen, C. N., and Burkell, J. 2002. "Believe It or Not: Factors Influencing C redibility on the Web," Journal of the American society for information science and technology (53:2), pp. 134 144. sefulness of Online Consumer Reviews," (17:1), pp. 19 38. Xu, Q. 2014. "Should I Trust Him? The Effects of Reviewer Profile Characteristics on Ewom Credibility," Computers in Human Behavior (33), pp. 136 144.

PAGE 99

. 88 Yin, D., Bond, S., and Zhang, H. 2014a. "Anxious or Angry? Effects of Discrete Emotions on the Perceived Helpfulness of Online Reviews," Mis Quarterly (38:2), pp. 539 560. Yin, G., Zhang, Q., and Li, Y. 2014b. "Effects of Emotional Valence and Arousal on C onsumer Perceptions of Online Review Helpfulness,"). Zadra, J. R., and Clore, G. L. 2011. "Emotion and Perception: The Role of Affective Information," Wiley interdisciplinary reviews: cognitive science (2:6), pp. 676 685. Zhang, J. Q., Craciun, G., and Shi n, D. 2010. "When Does Electronic Word of Mouth Matter? A Study of Consumer Product Reviews," Journal of Business Research (63:12), pp. 1336 1341. Zhang, W., and Watts, S. 2003. "Knowledge Adoption in Online Communities of Practice," ICIS 2003 Proceedings ) , p. 9. Zhang, W., and Watts, S. A. 2008. "Capitalizing on Content: Information Adoption in Two Online Communities," Journal of the Association for Information Systems (9:2), p. 73. Zhu, L., Yin, G., and He, W. 2014. "Is This Opinion Leader's Review Useful ? Peripheral Cues for Online Review Helpfulness," Journal of Electronic Commerce Research (15:4), p. 267.

PAGE 100

. 89 APPENDIX Appendix A: Summary of Review Helpfulness Studies Citation Independent Variables Method Sample Central cues Peripheral Cues ( Baek et al. 2012 ) Review length, Negative cords Rating inconsistency, Reviewer ranking, Reviewer real name Hierarchical regression Amazon N=75,226 ( Mudambi and Schuff 2010 ) Review depth (word count) Review extremity star rating Tobit Amazon N= 1587 ( Kuan et al. 2015 ) Review Leng th, Readability, Review valence Rating inconsistency, Reviewer credibility Probit, Heckman selection model Amazon N=89,362 ( Salehan and Kim 2014 ) Title sentiment, Title length, Review sentiment, Review length Review Longevity Negative binomial Amazon N=2616 ( Zhu et al. 2014 ) Review Readability, Review Length Reviewer expertise, Reviewer online attractiveness Negative binomial Yelp N=16,262 ( Huang et al. 2015 ) Word count Reviewer experience, Reviewer impact, Product Rating Tobit Amazon N=2209 ( Yin et al. 2014a ) Anxiety ( anxiety related words in a review) Ang er (anger related words in a review) Tobit Yahoo N= 7,322 ( Mousavizadeh et al. 2015 ) Utilitarian cues, Hedonic cues, Review Sentiment, Readability Review extremity, Review length, Title sentiment Negative binomial Amazon N=589

PAGE 101

. 90 Appendix B: Negative Binomial Regression Ordered Logit Variables Model Model Model Model Model 1 2 3 4 5 Elite Badge Member 0.35*** 0.34*** 0.34*** 0.34*** Review Rating 0.08*** 0.11*** 0.13*** 0.12*** Review Sentiment Polarity 0.13*** 0.07 0.04 0.04 Unsupervised Review Consistency 0.27*** 0.21** Supervised Review Consistency 0.30*** 0.1 Review Length 0.35*** 0.33*** 0.32*** 0.32*** 0.33*** Average Restaurant Rating 0.04 0.09 0.07 0.07 0.09 0.54*** 0.56*** 0.53*** 0.57*** 0.56*** Review Longevity 0.14*** 0.16*** 0.17*** 0.17*** 0.16*** Rating Inconsistency 0.26*** 0.25*** 0.23*** 0.23*** 0.23*** 0.07*** 0.03* 0.04* 0.04* 0.04* 0.10*** 0.09*** 0.09*** 0.09*** 0.09*** Tips 0.02* 0.01 0.01 0.01 0.01 0.01 0.004 0.01 0.01 0.01 Observation 2,234 2,234 2,234 2,234 2,234 Log Likelihood 2,624.67 2,605.51 2,589.19 2,590.85 2,605.51 Theta 4.43*** 4.73*** 5.12*** 5.17*** 4.73*** AIC 5,269.34 5,237.03 5,206.37 5,209.71 5,237.03