Expectation hurts? A dynamic model of the exogenous and endogenous role of eWOM and its effect on post-purchase responses
Student name: Xiaotong Chu student ID-card number: 11838515
Assignment: Master’s Thesis
Department: Graduate School of Communication
Program: Research Master of Communication Science Supervisor’s’ name: Guda van Noort
TABLE OF CONTENTS
ABSTRACT ... 3
INTRODUCTION ... 4
THEORETICAL BACKGROUND ... 6
EWOM dimensions: Valence, volume and quality ... 7
EWOM dimensions as exogenous variables: Driver of post-purchase responses and underlying mechanisms ... 9
EWOM dimensions as endogenous variables: The chronological evolution of eWOM dimensions ... 17
GENERAL METHOD ... 19
STUDY 1: EXPERIMENT ... 20
Method ... 20
Measures ... 22
Results ... 25
STUDY 2: AUTOMATIC CONTENT ANALYSIS ... 31
Method ... 31
Measures ... 32
Results ... 35
DISCUSSION ... 39
Theoretical implications ... 39
Limitations and directions for further research ... 42
REFERENCES ... 45
APPENDICES ... 54
Appendix II: Stimuli of experimentation ... 55 Appendix III: Online experimentation questionnaire ... 61 Appendix IV: Trigger word list of restaurants reviews ... 64
ABSTRACT
Building on the Expectation-Confirmation Theory (ECT), this study aims to: (1) investigate electronic word-of-mouth (eWOM) as an exogenous variable by comprehensively examining the effects of eWOM dimensions (valence, volume and quality) on post-purchase responses as well as the underlying cognitive mechanisms; (2) investigate eWOM as an endogenous variable predicted by exposure to previous eWOM; and (3) build a dynamic model to explain the temporal evolution of eWOM dimensions. A mixed-method approach, combining an experimental design and a large-scale automated content analysis, was adopted to examine the hypothesized model. Study 1 examined consumer responses evoked by eWOM dimensions in an experimentation and found that: (1) eWOM quality is the strongest predictor of consumers’ post-purchase satisfaction, attitude and repurchase intention, which can be explained by confirmation; (2) the positive effect of eWOM valence on post-purchase responses is counteracted with its negative effect on confirmation; and (3) the effect of eWOM volume on post-purchase responses is not supported. Study 2 examined the chronological evolution of eWOM dimensions in a large-scale data set of restaurant reviews and revealed that: (1) eWOM valence becomes more balanced over time, such that eWOM valences tends to decline when it starts high, while increase when it starts low; (2) eWOM volume encounters a steady increase until it reaches a peak and starts to decrease; and (3) eWOM quality tends to decline over time. Implications for academic and practical application of the dynamic model as well as limitations and directions for future research are discussed.
Keywords: eWOM, online reviews, Expectation-Confirmation Theory, expectation, confirmation, post-purchase responses, experimentation, automatic content analysis
INTRODUCTION
To date, the Internet has become an important venue for publicizing opinions and feedback. Therefore, electronic word-of-mouth (eWOM) is playing an increasingly important role in consumer affects, attitudes and purchase behaviors, which has drawn extensive attention from both the academics and practitioners. This study inspects the effects of eWOM from a multi-dimensional perspective. As previous studies suggested, the valence, volume and quality of eWOM each have a profound impact on consumer responses (Matute, Polo-Redondo & Utrillas, 2016; Kuo & Nakhata, 2019). So far, the valence and the volume of eWOM have received a lot of attentions (Cui, Lui & Guo, 2012; You, Vadakkepatt & Joshi, 2015), but the impact of eWOM quality has been largely disregarded. Therefore, the first aim of this study is to holistically explore the influence and evolution of eWOM valence, volume and quality.
With regard to the impact of eWOM, most studies focused on pre-purchase attitude and intention (Lascu, Bearden & Rose, 1995; Park, Lee & Han, 2007; Khare, Labrecque & Asare, 2011; Babić Rosario, Sotgiu, De Valck & Bijmolt, 2016). However, the impact of eWOM on post-purchase appraisals is lacking examination, which leads to a negligence of the role of eWOM in facilitating customer retention and preventing customer defection. Therefore, the second aim of this study is to shed lights on the multidimensional effects of eWOM on consumers’ post-purchase responses. To reach this aim, this study builds on the Expectation-Confirmation Theory (ECT), which posits that post-purchase evaluation is a function of a linear combination of expectations of the product performance and confirmation of expectation (Oliver, 1980; Bhattacherjee, 2001; Nam, Baker, Ahmad & Goo, 2019). However, the ECT treated expectation and confirmation both as exogenous variables and thus did not discuss the factors resulting in expectation and confirmation. This study adds eWOM to the ECT model, in order to investigate whether the eWOM dimensions influence variables
in the original ECT model.
EWOM serves not only as a precursor of consumers' pre- and post-purchase responses, but also as an outcome of consumers' post-purchase responses and decisions (Cheng & Zhou, 2010). Consumers refer to online reviews from previous consumers before the purchase decision, after which they also generate new online reviews based on their post-purchase evaluations. This circulated nature of eWOM determines that the eWOM effects should be examined in a dynamic manner. So far, a few studies have examined the evolution of eWOM valence (Melián-González, Bulchand-Gidumal & López-Valcárcel, 2013), while little attentions have been paid to how eWOM volume and quality evolve over time. Therefore, the third aim of this study is to develop and test a dynamic model of the endogenous role of eWOM with a focal point on the chronological evolution of eWOM dimensions.
To reach all the aims mentioned above, this study seeks to examine and answer the research question: to what extent do eWOM valence, volume and quality affect consumers’ post-purchase responses, which in return chronologically reflects on eWOM itself?
Due to the off-line nature of traditional word-of-mouth and inadequate computational technologies, previous academic approaches regarding eWOM effects were confined to experimental and field study settings (Oliver, 1980; Cheung & Thadani, 2012; Chong, Ch’ng, Liu & Li, 2017). In this research, the research question will be inspected with a combined methodology. First, an experimental study was conducted to examine the underlying cognitive processes mediating the effects of eWOM dimensions on consumers’ post-purchase affective, attitudinal and behavioral changes. Second, in order to provide comprehensive insights of eWOM as well as ameliorating the deficiency of the limited measurements of eWOM in the experiment, an automatic content analysis was conducted as an add-up approach to replenish the model with temporal effects. Large-scale data regarding the ratings, review texts and consumers’ visit time were collected from TripAdvisor.com to examine the
chronological evolution of eWOM dimensions. Machine learning was employed to generate insights in eWOM quality, which extends the current knowledge on eWOM quality measures.
In sum, we recognize the demand in the academic field of unifying and extending existing knowledge to accomplish a comprehensive conceptual model that addresses eWOM dynamics. Beyond previous literature concerning merely pre-purchase attitude and intention, we contribute the understandings of eWOM’s exogenous effects by providing insights in the multidimensional effects of eWOM on consumers' cognitive processes across different stages of purchase. Moreover, we examine the role of eWOM as an endogenous outcome of consumers’ post-purchase experience, which closes the loop of “eWOM – consumer responses - eWOM”. This study sheds light on the underexplored dynamic process of eWOM communication through an online experiment and with the help of an automatic content analysis as an additive and supplementary approach. A guideline is provided in order to help practitioners understand the temporal tendency of eWOM, improve brand image and consumer evaluation, as well as alleviating the impact of negative eWOM.
THEORETICAL BACKGROUND
To develop a dynamic model in which all three eWOM dimensions are inspected, three streams in the eWOM literature are drawn: (1) identifying the major dimensions of eWOM; (2) examining eWOM dimensions as exogenous variables that predict consumer post-purchase satisfaction, attitude and repurchase intention via expectation and confirmation; and (3) examining eWOM dimensions as endogenous variables predicted by consumers’ post-purchase responses and exposure to previous eWOM. Therefore, we demonstrate a cognitive model evoked by eWOM dimensions that exhibits explanation for the eWOM chronological evolution over time.
EWOM dimensions: Valence, volume and quality
As one of the oldest and most discussed mechanisms in the field of marketing, word-of-mouth refers to the act of consumers passing information and opinions on products or services to other consumers (Dellarocas, 2003; Babić Rosario et al., 2016). Word-of-mouth not only affects purchase behaviors, but also shapes consumers’ responses such as expectations, attitudes and behaviors (Arndt, 1967; Burzynski & Bayer, 1977; Zeithaml, Bitner & Dremler, 1996). With the emergence of the Internet, eWOM has become a substitution and supplement of traditional word-of-mouth, where online customer review serves as the most powerful channel (Dellarocas, 2003; Chevalier & Mayzlin, 2006). Statistics showed that eWOM has become an important source of information. According to Ecommerce Foundation’s 2018 Dutch Ecommerce Report1, 86% of Dutch consumers refer to online customer reviews, among which 50% read online reviews every time before purchasing, and 36% read online reviews occasionally.
Nowadays, eWOM has significantly overthrown the way that information used to be transmitted, and it transcended the limitations of traditional word-of-mouth communication (Duan, Gu & Whinston, 2008). Ellison and Fudenberg (1995) pointed out that as traditional word-of-mouth relies heavily on the social contact boundaries, its effect is limited over time and distance. Relative to conventional word-of-mouth communication, eWOM relies less on social tie strength, which indicates that eWOM is more influential for its low cost, high speed and worldwide network. However, eWOM also generates less credibility as it takes place between strangers instead of family and acquaintances (Park et al., 2007). Therefore, it is of especially importance in reassessing the diverse dimensional elements of word-of-mouth in a digital context as well as their respective effect.
Comparing to traditional word-of-mouth, eWOM interface of a product or service
normally carries three critical determinants of consumer responses (Park et al., 2007; Jiménez & Mendoza, 2013; Matute et al., 2016): (1) a rating score, or number of stars, of the product or service; (2) the number of reviews as an indicator of the population that has experienced the product or service; and (3) the review content which displays the reasons of review ratings. In this paper we will discuss and focus on these three main dimensions of eWOM: valence, volume and quality.
EWOM valence. Information varies in valence, which may be positive, neutral or
negative (Hodges 1974). This is also true for eWOM. The impact of different valence levels differs. Previous studies have indicated that comparing to positive eWOM communication, negative eWOM exerts much stronger impact on eWOM credibility (Xue & Zhou, 2010), consumer attitude toward the brand (Lee, Rodgers & Kim, 2009), attention to eWOM (Yang & Mai, 2010), consumer conformity (Lee, Park & Han, 2008), and product evaluation (Lee & Youn, 2009). Cheng and Zhou (2010) argued that negative eWOM has a greater influence because it reduces the possibility that the review is promoted by marketers, and thus a higher credibility (Cheng & Zhou, 2010).
EWOM volume. Another eWOM dimension that is most discussed in eWOM literature
is the amount of online reviews (Duan et al., 2008). Godes and Mayzlin (2004) argued that word-of-mouth volume indicates the popularity and awareness of a product or service, which is associated with the future rating. A higher eWOM volume leads to a more favorable attitudinal and behavioral responses (Park et al., 2007; Matute et al., 2016). The effect of eWOM volume tends to interact with the effect of eWOM valence, such that eWOM volume magnifies the effect of eWOM valence on consumer attitude and purchase intention (Tsao, Hsieh, Shih & Lin, 2015).
EWOM quality. One of the severest problems in the academic field of communication is
research, as studies on the effect of eWOM quality are relatively few (Matute et al., 2016). Moreover, findings are difficult to compare as the definition of eWOM content quality varies among studies. First, some scholars have argued that high-quality online review should present objective reasons when evaluating a product or service, while low-quality review simply displays subjective feelings (Park et al., 2007; Lee & Shin, 2014). Second, some other scholars argued that the diversity of topics contained in a review has an impact on consumer responses (Xu & Yao, 2015). Mitchell and Dacin (1996) indicated that a more comprehensive review is more likely to be written by a knowledgeable reviewer, which is translated into the review credibility. Therefore, an eWOM message that lacks diverse facts about the product or service is viewed as untrustworthy, while an eWOM message that contains various aspects is perceived to be credible (Jiménez & Mendoza, 2013). Third, review length (or word count) is also considered as a quality aspect, as it affects the perceived helpfulness of online reviews: it helps consumers to grasp more knowledge and features of the product or service (González-Rodríguez, Martínez-Torres & Toral, 2016). In addition, source expertise and homophily are also defined to be part of the eWOM quality (Cheng & Zhou, 2010), but here we do not take eWOM source into account as this study only focuses on the main text of the review in the examination of eWOM quality. In line with previous studies, we conceptualize eWOM quality along the lines of review length, objectivity and comprehensiveness.
EWOM dimensions as exogenous variables:
Driver of post-purchase responses and underlying mechanisms
Prior studies that assess the effect of eWOM have been largely focused on consumer attitudinal and behavioral responses preceding the purchase behavior (Fan & Miao, 2012; Erkan & Evans, 2016; Kudeshia & Kumar, 2017). This is a reflection of the industrial attention to acquire new audiences and grow a client base. Meanwhile, industries, such as
fast-moving consumer goods, financial services, tourism, not only seek to attract new customers, but also rely heavily on regular and frequent customers. Therefore, examination on eWOM effects on consumers’ responses during the post-purchase stage should attract equal attention as it helps with customer retention (Coyles & Gokey, 2005). Here we will explicitly clarify the effects of eWOM dimensions on post-purchase responses and the mediating cognitive processes based on the ECT.
EWOM effect on post-purchase responses. Consumer responses in the post-purchase
stage include affective, attitudinal and behavioral changes (Oliver, 1980; Hellier, Geursen, Carr & Rickard, 2003). As the first indicator of consumers’ post-purchase response, satisfaction is the psychological state based on the evaluation of the product or service during purchase experience, which is an emotional and affective dimension of the post-purchase responses (Oliver, 1981; Boulding, Kalra, Staelin & Zeithaml, 1993). Post-purchase attitude refers to a disposition or preference that consumers hold for the product or service, which is highly accepted by previous studies as an intermediate state between satisfaction and repurchase intention (Roest & Pieters, 1997; Hellier et al., 2003). In this study, these three variables are treated as indicators of post-purchase responses because they fully capture consumers’ the evaluation of the product and their experiences after the purchase behavior (Oliver, 1980; Park et al, 2007). According to Oliver (1980), post-purchase satisfaction, attitude and repurchase intention are highly correlated.
Previous studies have shown that different word-of-mouth dimensions have different effects on consumers’ post-purchase responses. With regard to eWOM valence, research suggested that positive eWOM positively induces post-purchase responses. For example, Burzynski and Bayer (1977) conducted a field study and the results demonstrated that consumers’ post-purchase attitude is altered by the valence of information cues, such that negative word-of-mouth (compared to positive word-of-mouth) leads to lower post-purchase
rating, and vice versa. Also, Kuo and Nakhata (2019) recently conducted three experiments and obtained similar results where positive eWOM results in higher degree of post-purchase satisfaction. Therefore, we speculate:
H1a: EWOM valence has a positive effect on post-purchase responses, such that eWOM with a more positive valence results in higher degrees of post-purchase satisfaction, attitude and repurchase intention.
With regard to the impact of eWOM volume, recent studies showed opposite results on repurchase intention. Matute et al. (2016) discovered that eWOM volume was a negative predictor of post-purchase attitude and repurchase intention, while Bulut and Karabulut (2018) found that eWOM volume has a positive total effect on repurchase intention through trust. As the later study did not confirm a direct effect of eWOM volume on post-purchase responses, we tend to formulate our hypothesis based on the former study. Matute et al. (2016) provided an explanation of the negative impact of eWOM volume on post-purchase responses: an overload of information could lead to confusion, which subsequently stops consumers from revisiting the online store for future purchases. Therefore, we postulate the following hypothesis:
H1b: EWOM volume has a negative effect on post-purchase responses, such that eWOM with a higher volume results in lower degrees of post-purchase satisfaction, attitude and repurchase intention.
As for the effect of eWOM quality, previous studies on the effects of eWOM have shown coherent results on consumer post-purchase responses. More specifically, the comprehensiveness is demonstrated to have a positive impact on post-purchase responses: the more diverse information provided in the message, the higher the post-purchase responses (Matute et al., 2016; Shin, Chung, Oh & Lee, 2013). Moreover, Bulut & Karabulut (2018) argued that a higher level of objectiveness is inclined to improve consumer post-purchase
satisfaction and repurchase intention, and thus helps to increase the likelihood of customer retention. The effect of another element of eWOM quality, the review length, has not been inspected in previous research, but due to its correlation with comprehensiveness (González-Rodríguez et al., 2016), a positive effect direction is expected. Therefore, we propose:
H1c: EWOM quality has a positive effect on post-purchase responses, such that eWOM with a higher quality results in higher degrees of post-purchase satisfaction, attitude and repurchase intention.
Underlying cognitive mechanisms explained by the ECT. We argue that the effect of
the three eWOM dimensions on consumers’ post-purchase reactions are mediated by cognitive processes, which refers to the underlying brain tasks that elicit actual affective, attitudinal and behavioral changes (Bandura, Adams & Beyer, 1977; Oliver, 1980; Schwenk, 1984). We identify two main lines in previous eWOM studies on the mediating cognitions on post-purchase responses: (1) information attribute-related variables such as perceived usefulness and credibility of the message (Matute et al., 2016; Bulut & Karabulut, 2018); and (2) consumer characteristics-related variables such as trust, involvement and loyalty (Praharjo, & Kusumawati, 2016). A third line regarding the mediating effects on post-purchase responses, yet lacking inspection in eWOM literature, is the expectation and confirmation proposed in the ECT model (Bhattacherjee, 2001). In this study, we focus only on the under-explored mediating effect of expectation and confirmation.
Widely employed to study consumer post-purchase responses, the ECT paradigm was developed by reviewing and coalescing two related theories, assimilation theory and contrast theory proposed by Sherif and Hovland (1961), which respectively provide perspectives of expectation (of perceived product performance) and confirmation (of expectation) (Oliver, 1977; Oliver, 1980). Oliver argued that post-purchase responses are influenced by the coupled
effect of expectation and confirmation, which constructs the main argument of the ECT. Oliver conducted two field experiments and found that both expectation and confirmation positively affect post-purchase responses, in the sense that post-purchase satisfaction, attitude and repurchase intention are inclined to coincide with the expectation and confirmation where the latter plays a greater role in magnitude. Here we respectively examine the mediating role of expectation and confirmation.
Mediating role of expectation. Expectation is the sum of consumers’ evaluation of
product performance, attitude and purchase intention preceding the purchase behavior (Oliver, 1980). Previous studies have pointed out that expectation that precedes purchase decision is shaped by word-of-mouth (Fazio & Zanna 1981; Bhattacherjee, 2001).
The effect of each eWOM dimension on pre-purchase attitude and behavior has been extensively examined. Park et al. (2007) conducted an experimental study on online shopping malls and found a positive relationship between pre-purchase attitude and intention and online reviews valence. Their findings are supported by other studies (Vermeulen & Seegers, 2009; Mauri & Minazzi, 2013). As expectation captures pre-purchase responses as a whole, we expect a positive effect of eWOM valence on expectation. Therefore, we hypothesize:
H2a: EWOM valence has a positive effect on expectation, such that eWOM with a more positive valence results in higher expectation.
As compared to valence, eWOM volume in fact provides a greater proxy of pre-purchase responses (Dellarocas, Zhang & Awad, 2007; Khare et al., 2011; Babić Rosario et al., 2016). This can be explained by consumer conformity theory: individuals tend to generate favorable attitudes toward an opinion of a group of referents (Lascu et al., 1995; Park et al., 2007). In a similar vein, we expect a positive effect of eWOM volume on expectation. Therefore, we speculate:
higher volume results in higher expectation.
Qualitative characteristics of word-of-mouth content negatively affect attitude of the message receiver, particularly in the online context. Jeong and Koo (2015) conducted an online experiment and found that consumers’ attitude and purchase intention toward the product or service is negatively related to eWOM objectivity. Filieri (2015) argued that those exposed to lengthy reviews with more varying and detailed information are inclined to evaluate the product or service more positively. Same findings were supported by the experimental study conducted by Lee and Shin (2014). Accordingly, we expect a positive effect of eWOM quality on expectation, and the hypothesis is as following:
H2c. EWOM quality has a negative significant effect on expectation, such that eWOM with a higher quality results in lower expectation.
Expectation provides the baseline for post-purchase responses, such that a higher expectation enhances consumers’ post-purchase attitude and intention, while a lower expectation reduces consequent responses (Oliver, 1980; Bhattacherjee, 2001). Boulding et al. (1993) conducted a longitudinal laboratory experiment where satisfaction is once again confirmed as a function of prior expectation, which predicts consumers’ attitude and behavioral intention. Thus:
H2d. Expectation has a positive effect on post-purchase satisfaction, attitude and repurchase intention.
Mediating role of confirmation. Expectation is not the only standard consumers use
when evaluating their purchase experience (Cadotte, Woodruff & Jenkins, 1987). As the most immediate cognitive process after purchase behavior, confirmation refers to the extent to which the product or service exceeds, meets or eludes consumers’ expectations antecedent to purchase (Oliver, 1980; Bhattacherjee, 2001). Viewed as a deviation from the expectation level, confirmation varies on an indicator from “negatively disconfirmed” to “confirmed” to
“positively disconfirmed” (see Figure 1). A disconfirmation happens when perceived product performance fails or exceeds expectation, while a confirmation happens when expectation is met (Oliver, 1977). Although the explanatory power of confirmation is demonstrated to be stronger than that of expectation (Oliver, 1980; Bhattacherjee, 2001), the effect of confirmation has been greatly neglected in previous studies.
Figure 1
Magnitude and direction of confirmation
Given that the social tie in eWOM is relatively weak, the valence and volume of eWOM, indicating the favorability and popularity of the product or service, tend to be viewed as less trustworthy and less persuasive (Park et al., 2007; Duan et al., 2008). Wood and Eagly (1981) conducted an experimentation and discovered that consumers tend to encounter an disconfirmation when a message is potentially biased. Therefore, consumers are inclined to generate confirmation inversely, in the sense that online reviews with more positive valence and higher volume lead to negative disconfirmation, while online reviews with more negative valence and lower volume result in positive disconfirmation. We derive the hypotheses as follows:
H3a. EWOM valence has a negative effect on confirmation, such that eWOM with a more positive valence results in lower confirmation.
H3b. EWOM volume has a negative effect on confirmation, such that eWOM with a higher volume results in lower confirmation.
EWOM quality indicates the accuracy and persuasive strength in online reviews, which can be captured by confirmation (Nam et al., 2019). In other words, when referring to eWOM for product evaluation, consumers tend to evaluate the credibility and usefulness of the
message, such that online reviews with higher level of objectivity and comprehensiveness enhance consumers’ confirmation of expectation, while subjective online reviews with insufficient information leads to negative disconfirmation of expectation. Therefore, we presume that:
H3c. EWOM quality has a positive effect on confirmation, such that eWOM with a higher quality results in higher confirmation.
As confirmation is defined as the discrepancy between expectation and product performance, the relationship between expectation and confirmation depends on the manipulation of expectation and perceived product performance (Oliver, 1977): (1) when viewing expectation and performance as exogenous variables, expectation and confirmation exist independently (Oliver, 1977; Oliver, 1980); (2) when viewing confirmation as a cognitive appraisal of expectation and manipulating expectation, confirmation is negatively correlated with expectation (Bhattacherjee, 2001; Kim, Ferrin & Rao, 2009); (3) when viewing confirmation as a performance-based appraisal and manipulating perceived product performance, confirmation is positively correlated with expectation (Kim, 2012). In this study, we argue that expectation is influence by eWOM, thus we speculate as follows:
H3d. Expectation has a negative effect on confirmation.
H3e. Expectation mediates the effects of eWOM valence, volume and quality on confirmation.
The formation of post-purchase attitude and intentions is not only assimilated toward expectation, but also largely predicted by confirmation. Customers form their post-purchase attitude by comparing their perceptions of the product or service before and after purchase (Ryu & Han, 2010; Chen, Yen & Hwang, 2012). Nam, Baker, Ahmad & Goo (2018) conducted a survey study with data collected from users of the TripAdvisor.com and found that eWOM-based confirmation has a positive effect on post-purchase experiences. Moreover,
Oliver (1980) addressed that the effect of confirmation on post-purchase experiences is greater in magnitude than that of expectation. Hence, the hypothesis is as following:
H3f. Confirmation has a positive effect on post-purchase satisfaction, attitude and repurchase intention, and this effect is stronger than the effect of expectation.
EWOM dimensions as endogenous variables: The chronological evolution of eWOM dimensions
Although researchers have realized eWOM’s role as an outcome of consumers’ responses, eWOM has been seldom investigated as an endogenous variable. To accomplish the dynamic model and close the eWOM cycle, this study proposes the chronological evolution of eWOM dimensions based on consumers’ post-purchase responses. More specifically, consumers’ purchase experience results in generating eWOM, which adds to the existing set of eWOM dimensions for new consumers who subsequently refer to the new set of eWOM and generate pre- and post-purchase responses.
Chronological evolution of eWOM valence. The formation of eWOM valence is
significantly in line with customer post-purchase responses (Suki, 2014). As we discussed, eWOM valence has a negative effect on confirmation, which subsequently has a greater explanatory power on post-purchase evaluations. Previous studies have supported our assumption. On the one hand, Li and Hitt (2008) conducted a study on book sales on Amazon.com, where the results suggest that eWOM at an early stage demonstrates a positive bias, and that consumer ratings decline over time. On the other hand, Melián-González et al. (2013) investigated hotel reviews on TripAdvisor.com, where they found that early eWOM tend to be negative, but as the negative effect is subsequently mitigated over time. Hu, Liu and Zhang (2008) found a relationship between eWOM valence and time where the longer a product or service exists, the smaller the effect of eWOM valence is, thus eWOM valence
tends to approach to a balanced level over time. This finding was supported by Babić Rosario et al. (2016) where they found that the connection with eWOM is stronger for new products than for mature products. Combing the findings of previous studies, we hypothesize:
H4a. Positive early eWOM encounters a decrease in the valence over time, while negative early eWOM encounters an increase in the valence over time; in the end, eWOM valence tends to approach to an averaged level.
Chronological evolution of eWOM volume. The growth of the Internet allows more
people to publish opinions and evaluations, which inevitably generates a higher eWOM volume (Matute et al., 2016). However, when the eWOM volume reaches a certain level, the rate of volume increase tends to decrease. Hu et al. (2008) posited a relationship between eWOM volume and time, such that as the age of a product or service grows, eWOM volume increases at a decreasing rate. Therefore, we presume:
H4b. EWOM volume increases at an increasing rate at an early stage, and it increases at a decreasing rate after the rate reaches the peak.
Chronological evolution of eWOM quality. At an early stage, the qualitative
characteristic of eWOM is strengthened as prior eWOM provides informative and objective insights. Hu et al. (2008) argued that consumers are more motivated to generate online reviews when the coverage of the product or service is low, thus new online reviews tend to be more comprehensive than formerly created reviews. As the coverage of eWOM has reached a certain level, generating new perspectives becomes challenging, thus generating qualitative characteristics of eWOM is discouraged. Moreover, the growth of eWOM volume dilutes the proportion of online reviews written by knowledgeable buyers and subsequently diminishes the eWOM quality (Matute et al., 2016). Therefore, we expect:
H4c. EWOM quality increases at an early stage, and then decreases after it reaches the peak.
After adding eWOM at different time points in the model, the dynamic model is closed and formed into a cycle. Figure 2 provides an overview of the proposed model as well as the assumptions. The proposed model indicates that eWOM dimensions affect post-purchase responses through expectation and confirmation, which subsequently has an effect on the newly generated eWOM, which will add to the existing set of eWOM and subsequently affect other consumers’ responses.
Figure 2
Proposed dynamic conceptual model of eWOM dimensions, cognitive processes, and post-purchase responses
GENERAL METHOD
Study 1 applied experimentation to test the sub-hypotheses of H1, H2 and H3, where we investigated the effects of eWOM valence, volume and quality on post-purchase responses as well as the underlying cognitive mechanisms. Study 2 applied automatic content analysis to test the sub-hypotheses of H4, where we inspected the real-world chronological evolution of eWOM valence, volume and quality.
STUDY 1: EXPERIMENT Method
Design. 3 (eWOM valence: positive vs. negative vs. mixed reviews) × 2 (eWOM
volume: high vs. low) × 2 (eWOM quality: high vs. low) between-subjects design was adopted to test the hypotheses.
Participants. Participants (N = 389) were recruited via Amazon Mechanical Turk, and
each participant received 1 USD for participating in the study. The average age of participants was 34.61 (SD = 9.27), and 63.0% of them were male. With regard to the educational level, 59.6% had received a bachelor’s degree, 18.3% had received a high school or equivalent degree, and 10.0% had received a master’s degree.
Experimental material. The stimuli consisted of three components, namely a purchase
scenario, online review section, and the actual purchase experience of playing an online game. Participants were firstly instructed to a purchase scenario where they were asked to imagine that they were about to purchase an online game product. Next, participants were exposed to the review section, and then they were asked to play the game as an actual purchase behavior. Only the online review section differed between conditions, while the purchase scenario and game element were held constant.
The purchase of an online game was chosen as the experimental context because of two reasons: (1) the nature of online gaming determines that it can only be purchased in an online context; (2) online game consumers are inclined to refer to reviews from other players (Zagal, Ladd & Johnson, 2009). The World Hardest Game (see Appendix I) was selected due to its polarized feedbacks and a relatively small number of reviews (by 2019, December 31, this game has received 197 reviews with an average rating of 3.2 out of 5 in Microsoft online
store2). Therefore, the likelihoods of product performance and existing attitudes having impacts on the outcomes were lowered.
After the purchase scenario, participants were exposed to the online reviews of the game. This section differed across conditions. To increase external validity, the layout of the review section was borrowed from Microsoft online store. Figure 3 shows an example of the stimuli (all the other eWOM stimuli are in Appendix II).
Figure 3
Online review example
EWOM valence was reflected in the rating score on the top left of the review section. Based on Hartman, Hunt and Childers’s experimental study (2013), the eWOM valence variation was divided into three conditions: positive rating (4.8 out of 5) with two positive reviews, negative rating (1.2 out of 5) with two negative reviews, and mixed rating (3 out of 5) with a positive review and a negative review. The number of reviews reflected the eWOM volume, which was set to either 2000 (high volume) or 10 (low volume) based on the number
2 The link of the game in Microsoft online store:
of reviews that the most and least popular games received in Microsoft online store. EWOM quality was reflected in the content of online reviews, which was extracted from the existing reviews of this game in order to avoid misleading information. A high quality online review contains five long sentences with logical, objective and diverse evaluations of the game; a low quality online review contains three short sentences that contained a slang, subjective, emotional and superlative words, and it was lack of detailed information regarding the game (Park et al., 2007).
Procedures. The experiment was administered online via MTurk. After signing the
informed consent letter, participants were told to imagine purchasing an online game product. Participants were then randomly assigned to one of the twelve conditions where they were asked to read the ratings and reviews of an online game The World Hardest Game, as well as answering questions regarding expectation. Afterwards, participants were asked to click a hyperlink3, which lead to the official website of the game, where they were asked to play the game and try to pass the first level. After playing the game, questions regarding confirmation and post-purchase responses were asked. Next, control variables were measured regarding the online reviews and game experiences. In the end, manipulation check questions and demographic questions (such as age, gender and education level) were asked. Finally, respondents were debriefed and thanked for their participation.
Measures
Mediators (expectation and confirmation), dependent variables (post-purchase satisfaction, attitude and repurchase intention) and control variables were measured in the experiment (see Appendix III). For all the construct measures, existing and validated scales were used. Each construct has a Cronbach's alpha score of .80 or higher (see Table 1 for
descriptive statistics of each scale). Table 1
Overview of the measures and descriptive statistics
Mediators. Expectation was measured with participants’ perceptions of the video
game’s overall quality by answering the question: “Before you play the game, do you expect the game to be …?” on a 7-point Bipolar scale from “very poor” to “very good” (Oliver, 1977). Besides, the expected product performance was measured with three items regarding online gaming attributes (graphics, sound, and controls) on a 7-point Likert scale from “very poor” to “very good” (Oliver, 1977; Yang & Mai, 2010). These four items together had an adequate internal consistency.
Confirmation refers to the discrepancy between expectation and actual product performance, which can be measured in two ways: the overall expectancy confirmation, and the changes in the attribute ratings between pre- and post-purchase experiences (Oliver, 1977; Bhattacherjee, 2001). The overall confirmation was measured by the item “after playing the game, I think the game is …” on a 7-point scale from “poorer than I expected” (indicating that expectation was negatively disconfirmed) to “exactly as I had expected”(indicating that expectation was confirmed), to “better than I expected” (indicating that expectation was positively disconfirmed (Oliver, 1977). In order to obtain an attribute rating difference score, ratings of game attributes (graphics, sound, and controls) were measured again after the
exposure to the game on a 7-point Likert scale from “very poor” to “very good”. Differences between the attribute ratings before and after playing the game were calculated where a positive (or negative) score indicated a positive (or negative) expectation disconfirmation, and a score of zero indicated expectation confirmation. These four items together had an adequate internal consistency.
Dependent variables. As an overall emotional response to the purchase experience,
post-purchase satisfaction was measured with two items (“I am satisfied with the game”, and “I feel delighted with the game”) on a 7-point Likert scale from “strongly disagree” to “strongly agree” (Spreng & Mackoy, 1996; Nam, Ekinci & Whyatt, 2011).
Post-purchase attitude toward the game was measured on four 7-point Bipolar scales: from “unappealing” to “appealing”, from “unpleasant” to “pleasant”, from “unfavorable/” to “favorable”, and from “unlikable” to “likable” (Spears & Singh, 2004). Repurchase intention was measured with the likelihood of repurchasing the game and recommending the game to friends on two 7-point Bipolar scales from “very unlikely” to “very likely” (Jamieson & Bass, 1989; Kalwani & Silk, 1982; Hartmann & Apaolaza-Ibáñez, 2012; Park et al., 2007).
Control variables. Participants were asked to answer questions about their general
attitudes toward online reviews and their experiences with the game. The prior usage of online reviews was measured by one item “I always read online reviews before I buy a product” on a 7-point Likert Scale from “strongly disagree” to “strongly agree”, while the general attitude toward online reviews was measured by two items (perceived usefulness and trust) on a 7-point Likert Scale from “strongly disagree” to “strongly agree” (Park et al., 2007). These three items together had an adequate internal consistency (M = 5.76, SD = 0.89, Cronbach’s a = .79).
As purchase experiences have an impact on post-purchase responses (Sullivan & Kim, 2018), variables regarding the exposure to the online game were measured. Time devoted on
the game during the experiment was asked because it is related to participants’ evaluation and preference of a game (Yang & Mai, 2010). 49.1% participants spent less than 5 minutes on the game, while 42.9% participants spent five to ten minutes on the game. Only 1.3% participants played the game for more than 15 minutes. Prior experience with the game was measured by asking participants if they had played the game before the experiment in order to detect the presence of preconceived attitudes, and 15.7% of the participants had played the game before.
Results
Randomization check. Results of the randomization check revealed that experimental
conditions did not significantly differ among each other regarding age (F(11,377) = 0.84, p = .604), gender (X²(11) = 7.25 p = .778), education level (X²(11) = 39.31 p = .945; Likelihood Ratio²(33) = 37.80, p = .963), attitudes toward online reviews (F(11,377) = 0.94, p = .500), the time spent on playing the game during the experiment (X²(33) = 22.11 p = .925; Likelihood Ratio²(33) = 24.11, p = .870), and prior-knowledge of the game (X²(11) = 14.27 p
= .219). Therefore, demographic characteristics and control variables are randomly distributed among the experimental conditions.
Manipulation check. The efficacy of manipulations was inspected with 9-point numeric
scales where a higher score means that participants perceived a more positive valence, a higher volume or a higher quality (Khare et al., 2011).
The eWOM valence manipulation was assessed with two 9-point items: “The reviewers’ evaluations of the game were …” (very negative/ very positive), and “The reviewers’ opinions about the game were …” (very unfavorable/ very favorable; M = 5.80, SD = 2.74, Cronbach’s a = .97). The results of an analysis of variance (ANOVA) indicated that the eWOM valence conditions significantly differed from each other in the expected direction
(MPositive_valence = 7.91, SD= 1.22; MNegative_valence = 3.45, SD= 2.90; MMixed_valence = 5.95, SD=
1.73; F(2, 386) = 151.45, p < .01).
To test the manipulation of eWOM volume, two 9-point items were employed: “The game was played by …” (very few people/ very many people), and “The opinions expressed about the game were …” (very few/ very many; M = 5.40, SD = 2.49, Cronbach’s a = .90). An independent t-test was conducted and demonstrated that the eWOM volume conditions differed significantly (MHigh_volume = 5.99, SD= 2.38; MLow_volum = 4.75, SD= 2.44; t(387) =
5.05, p < .01).
To assess the efficacy of the eWOM quality manipulation, four items were measured on a 9-point scale: “The reviews of the game were …” (very short/ very long), “The reviews of the game were …” from (very subjective/ very objective), “The reasons supporting reviewers’ opinions were ...” (very insufficient/ very sufficient), and “The credibility of the reviews was …” (very low/ very high) (M = 5.49, SD = 1.92, Cronbach’s a = .87). An independent t-test was conducted and demonstrated that the eWOM quality conditions significantly differed in the expected direction (MHigher_quality = 5.85, SD= 1.59 MLower_quality = 5.12, SD=
2.15; t(387) = 3.80, p < .01).
Correlations between variables. Pearson correlations between variables were reported
(see Table 2), so as to get an overview of all the measured variables, as well as screening out a multi-collinearity problem when testing hypotheses using MANOVA (Meyers, Gamst & Guarino, 2016). The analysis showed significant correlations among all the variables. Specifically, dependent variables were moderately correlated. Mediating variables expectation and confirmation were negatively correlated, while both the expectation and the confirmation were positively related to dependent variables (post-purchase satisfaction, attitude and repurchase intention). Based on the significant correlations between variables, the magnitudes and directions of effects among mediating and dependent variables can be expected.
Table 2
Pearson correlations between mediating and dependent variables
Testing hypotheses. Hypotheses were tested in the following paragraphs. In specific, the
direct effects of eWOM dimensions on post-purchase responses (proposed in H1a, H1b and H1c) were tested with a MANOVA. Next, the effects of eWOM dimensions on expectation (proposed in H2a, H2b and H2c) and the direct effects of eWOM dimensions on confirmation (proposed in H3a, H3b and H3c) were respectively tested with a two-way ANOVA. The effect of expectation on confirmation (proposed in H3d) was examined with a linear regression analysis, while the mediating role of expectation on the effect of eWOM on confirmation (proposed in H3e) was tested with Hayes’ PROCESS Macro. In addition, the effect of (proposed in H2d and H3f) was examined with a multiple regression analysis.
Effects of eWOM dimensions on post-purchase responses. Sub-hypotheses of H1
hypothesized the effects of three independent variables (eWOM valence, volume and quality) on three dependent variables (post-purchase satisfaction, attitude and repurchase intention), which were tested by performing a combined factorial MANOVA where the usage and attitude toward online reviews as well as the time spent and prior knowledge of the game were treated as covariate variables. As shown in Table 3, neither eWOM valence nor volume had significant effect on post-purchase satisfaction, attitude and repurchase intention, thus H1a and H1b were not supported. However, eWOM quality had a significant positive effect on post-purchase satisfaction (Mhigher_quality = 4.64, SD= 0.12; MLower_quality = 4.13, SD= 0.12),
attitude (Mhigher_quality = 4.77, SD = 0.13; MLower_quality = 4.32, SD= 0.13), and repurchase
intention (Mhigher_quality = 4.50, SD= 0.15; MLower_quality = 4.03, SD= 0.14), Wilks’ Lambda
= .977, F (3, 372) = 2.95, p = .033. Therefore, compared to online reviews with lower quality, online reviews with higher quality results in higher degrees of post-purchase responses. Thus H1c was supported.
Table 3
Factorial MANOVA results of eWOM effects on post-purchase satisfaction, attitude and repurchase intention
Effects of eWOM dimensions on expectation and confirmation. Sub-Hypotheses of H2
assumed the impact of three independent variables (eWOM valence, volume and quality) on expectation, which were test by a two-way ANOVA where the usage and attitude toward online reviews were treated as covariate variables. As shown in Table 4, eWOM valence had a significant effect on expectation (MPositive_valence = 5.97, SD= 0.13; MNegative_valence = 3.10, SD
= 0.13; MMixed_valence = 4.89, SD= 0.13), which indicated that more positive eWOM resulted in
expectation level was significantly different among positive, negative and mixed eWOM conditions (p = .000 across all three groups). Therefore, H2a was supported. The same analyses with eWOM volume and quality as factors demonstrated that neither eWOM volume nor eWOM quality had a significant effect on expectation, thus H2b and H2c were not supported.
Table 4
ANOVA results of eWOM effects on expectation and confirmation
H3a, H3b and H3c predicted the effects of three independent variables (eWOM valence, volume and quality) on confirmation. Table 4 displays the results of a two-way ANOVA where the usage and attitude toward eWOM as well as game experience were treated as covariate variables. The results revealed a significantly negative impact of eWOM valence on confirmation (MPositive_valence = 0.15, SD= 0.12; MNegative_valence = 1.96, SD= 0.12; MMixed_valence
= 1.00, SD= 0.12). This finding showed that exposure to negative eWOM leads to a more positive disconfirmation, while exposure to positive eWOM leads to a more negative disconfirmation. Pairwise comparisons of marginal means was conducted and revealed that
confirmation level was significantly different among positive, negative and mixed eWOM conditions (p = .000 across all three groups). Therefore, H3a was supported. EWOM volume had no significant effect on confirmation, thus H3b was not supported. EWOM quality was observed to have a significant effect on confirmation (Mhigher_quality = 1.30, SD = 0.10;
MLower_quality = 0.80, SD= 0.10), which indicated that exposure to eWOM of higher quality,
comparing to that of lower quality, resulted in a more positive disconfirmation. Therefore, H3c was supported.
A linear regression analysis was conducted to test the effect of expectation on confirmation hypothesized in H3a (b = -.53, t(387)= -11.66, p = .000). A significant regression equation was found (F(1,387) = 39.52, p = .000), with an R2 of .29. The result demonstrated that expectation had a negative effect on confirmation, thus H3d was supported.
Based on the findings above, we speculate that expectation mediates the effect of eWOM valence on confirmation. In order to test H3e, we employed the PROCESS Macro for SPSS with 5,000 bootstrap sample (Hayes, 2017). The results showed that the indirect effect of eWOM on confirmation was significantly greater than zero (BootLLCInegative_valence = 0.60,
BootULCInegative_valence = 1.21; BootLLCImixed_valence = 0.21, BootULCImixed_valence = 0.50).
Therefore, we concluded that expectation fully mediated the relationship between eWOM valence and confirmation, and H4e was supported regarding to the effect of eWOM valence.
Effects of expectation and confirmation on post-purchase responses. Multiple
regression analyses were conducted to examine the effects of two mediating variables (expectation and confirmation) on three dependent variables (post-purchase satisfaction, attitude and intention). The results showed that both the expectation and confirmation were significantly positive antecedent of post-purchase satisfaction, attitude and repurchase intention, where confirmation had a greater explanatory power (see Table 5). The findings were consistent with H2d and H3f.
Table 5
Multiple regression results of the effects of expectation and confirmation on post-purchase satisfaction, attitude and intention
Findings in Study 1 examined eWOM dimensions as an input of post-purchase responses, as well as underscoring the especially important mediating role of expectation and confirmation. Based on which, Study 2 was conducted to extend findings in Study 1 by inspecting eWOM as an output of post-purchase responses, which aims to close the dynamic loop of eWOM effects. The dual purposes of Study 2 were to study the chronological evolution of eWOM dimensions in real world.
STUDY 2: AUTOMATIC CONTENT ANALYSIS Method
Sample. Restaurants on TripAdvisor.com was selected for this study due to three reasons:
(1) comparing to merchandise industry, restaurant as a service industry replies heavily on customer retention, thus the eWOM effect is amplified; (2) while eWOM on e-commerce sites such as Amazon were fully studied during the past years (Hu et al., 2008; Mudambi & Schuff, 2010), travelling sites as another important eWOM community that consumers tend to refer to was lack of examination; (3) as the world's largest travel site, TripAdvisor.com has a
well-structured online review and rating system (Melián-González et al., 2013). According to the 2019 TripAdvisor Review Transparency Report4, TripAdvisor.com received 66 million online reviews in 2018, among which 53% of the reviews were for locations in Europe. The average rating submitted by reviewers was 4.22 out of 5. Amsterdam, the capital of the Netherlands, was selected to draw a sample due to two reasons: first, more than half of the users on TripAdvisor.com located in the Europe, and Amsterdam was one of the largest and most international European cities; second, given our focus on English online reviews, the fact that the Dutch speak fluent English enabled consumers to read and write English online reviews (Hardingham-Gill, 2019).
Data collection. Information from TripAdvisor.com was obtained through webpage
scraper, which was written in Python 3. The procedure of data collection started on 2019, December 15, and lasted for five days. There were 3,796 restaurants in Amsterdam published on TripAdvisor.com, among which 3,562 restaurants had recorded ratings and reviews. Due to a high correlation between the number of English reviews and the number of reviews of all the languages (r = .98, n = 3562, p < .01), hypotheses were tested using merely the number of English reviews. In total, we collected 301,996 English reviews from 3,562 restaurants within an accessible time range between 2010 and 2019. The average rating per restaurant was 4.16 (SD = 0.68), among which 82.03% were ratings of 4 or 5, while 9.04% were ratings of 1 or 2.
Measures
Age of restaurant. To inspect the dynamics of eWOM dimensions, the temporal
properties of online reviews is defined. Hu et al. (2008) introduced the concept “age of item”, which refers to the time range from the date when an item was published online till the date when the data was collected. Taking into account that seasonal changes play a particular role
in business fluctuation in the restaurant industry, online reviews were grouped by the age of the respective restaurant in years. For example, a restaurant at an age of seven years refers to the seventh year since the restaurant started receiving online reviews on TripAdvisor.com (see Table 6). 823 out of 3,562 restaurants had an age of eight years or more, among which 747 restaurants received online reviews across all the eight age years (185,127 reviews).
Table 6
Age of restaurant
EWOM valence. On TripAdvisor.com, reviewers are asked to give an integer number of
rating from 1 to 5 when they leave a review. The review site categorizes theses ratings in five categories: 1 (terrible), 2 (poor), 3 (average), 4 (very good), and 5 (excellent). To test the temporal effects of eWOM valence in a straightforward manner, we calculated the averaged cumulative ratings for each restaurant at each age (Melián-González et al., 2013).
EWOM volume. EWOM volume was indicated by the number of English reviews a
restaurant received on TripAdvisor.com in each age.
EWOM quality. The eWOM quality score was based on three elements in each review:
the length, objectivity and comprehensiveness. The score of each element ranged from 0 to 1 where a lower value indicated lower length, lower objectivity (higher subjectivity), and less diverse information provided. Each score was respectively calculated for each review, and then an averaged score of eWOM quality was calculated on a restaurant level at each age.
which ranged from 1 to 2131 words (M = 70.45, SD = 60.12). 50% of the reviews fell within an interval from 33 to 87 words, while 90% of the reviews fell within an interval from 20 to 179 words. In order to obtain a value within an interval from 0 (very short) to 1 (very long), the value of online review length was normalized with the following equation (Jain & Bhandare, 2011):
!"#$%&'()_!"#$% = !"#$%!− min (!"#$%) max !"#!" − min (!"#$%)5
Second, the level of objectivity for each review was obtained with Vader lexicon (Hutto & Gilbert, 2014; Ahmad & Laroche, 2017), which returned a value from 0 (very subjective) to 1 (very objective). With an accuracy score of 0.96 according to Singh and Vasudeva’s study (2018) on online restaurant reviews, Vader returns a dictionary with four scores: positivity, negativity, neutrality, and compound. In this study, the “neutral” score was employed as the objectivity level of online reviews (M = .73, SD = .12).
Third, the comprehensiveness score was generated by aspect extraction using a dictionary-based approach (Humphreys & Wang, 2018). A Convolutional Neural Network (CNN) model was employed for tagging trigger words in online reviews into six different aspects: food and drink, service, atmosphere, value, location and overall (LeCun & Bengio, 1995; Poria, Chaturvedi, Cambria & Bisio, 2016). The dictionary of trigger words was constructed using a pre-trained CNN model especially designed for extracting restaurant review aspects6 (; Xu, Liu, Shu & Yu, 2018). To get a list of trigger words that grasped the meaning of the review text, we randomly selected 5,000 online reviews from the dataset. Subsequently, we obtained a list of 4,430 trigger words which was sorted by word occurrences, and we selected the top 10% words from this list that has an occurrence of 10 times or higher. Afterwards, we manually assigned each trigger word into one of the six
aspects and created a dictionary with aspects (see Appendix IV). We then parsed the review texts using this dictionary, which started with extracting textual data from each online review that resulted in a matrix of words. Data preprocessing was conducted, which included linguistic reduction by removing punctuations, as well as converting characters to lowercases. As each trigger word indicated the presence of an aspect in an online review, we calculated the comprehensiveness score based on the number discrete aspects were extracted by the trigger words (see Figure 4). The comprehensiveness score was normalized to a range from 0 to 1 where lower scores indicated less diverse information in online reviews.
Figure 4
Example of generating comprehensiveness score of review texts
Results
Chronological evolution of eWOM valence. H4a assumed that eWOM valence tends to
approach to a balanced level over time, in the sense that eWOM valence decreases when it starts high, and increases when it starts low. Based on the average rating of each restaurant in the first age, median split was applied to divide the sample into two groups (Sher & Lee,
2009): restaurants with an averaged rating of 4.2 or higher in the first age year were assigned to the higher early rating group (n = 375), while restaurants with an averaged rating lower than 4.2 in the first age year were assigned to the lower early rating group (n = 372).
Dependent t-test analyses were applied to analyze the rating differences among ages. Table 7 reveals that restaurants in the higher early rating group encountered a significant decrease in their ratings in subsequent age years, and then the decrease slowed down as it approached to the overall average rating; restaurants in the lower early rating group encountered a significant increase in their ratings in subsequent years, and the growth slowed down as it approached to the overall average rating. In addition, the standard deviation of ratings in the higher early rating group slightly increased over time, while that in the lower early rating group declined overtime, which indicated a greater polarization in higher early rating group and a smaller polarization in lower early rating group. In general, restaurant ratings did not experience a significant change, and standard deviation indicated a greater agreement in ratings overall. Therefore, H4a was supported. The chronological evolution of eWOM valence over time is displayed in Figure 5.
Table 7
Figure 5
Chronological evolution of the averaged cumulative ratings of restaurants
Chronological evolution of eWOM volume. H4b presumed that eWOM volume
increases at an increasing rate at first, and then increases at a decreasing rate after reaching the peak. To examine the growth rate of eWOM volume, we calculated the averaged number of online reviews that restaurants received in each age year. Dependent t-test analyses were applied to test the volume differences over time. Table 8 reveals that the number of online reviews significantly increased in the first six age years, and then declined after reaching the peak. Therefore, H4b was supported. EWOM volume evolution is displayed in Figure 6. Table 8
Figure 6
Chronological evolution of the number of online reviews restaurants received
Chronological evolution of eWOM quality. H4c hypothesized that eWOM quality
increases at first, and then decreases after reaching the peak. Dependent t-test analyses were applied to analyze the eWOM quality evolution over time. According to Table 9 and Figure 4, all the three elements in eWOM quality encountered a decrease over time: (1) review length significantly declined in the first five age years and later leveled off; (2) except for the second age year, objectivity of online reviews significantly decreased in the first five age years and later flattened; (3) comprehensiveness of online reviews encountered a significant decrease from the second to the fourth age year. Therefore, H4c was not supported.
Table 9
Figure 7
Chronological evolution of online review quality
DISCUSSION
This study aims to extend current research on eWOM by applying and extending the ECT model to explain the effects of eWOM valence, volume and quality on consumers’ post-purchase responses. As a consequence, we developed a closed dynamic model of cognitive processes evoked by eWOM. The model was tested with a combined approach of experimentation and automatic content analysis, and interesting findings were addressed.
Theoretical implications
This study calls attention to three under-investigated perspectives of eWOM research. First, among the limited amount of studies focusing on consumer responses in the post-purchase stage, few have simultaneously inspected the impacts of all the three dimensions of eWOM: valence, volume and quality. Meanwhile, mixed findings are present in literatures. In Study 1 where an experimental study was conducted, we demonstrate that
different eWOM dimensions have different cognitive patterns of impact on post-purchase affective, attitudinal and behavioral consequences: eWOM quality has a positive impact on post-purchase responses, while eWOM valence and volume have no significant impact on post-purchase responses.
Second, this study applies the ECT model to eWOM effects, which helps to extend the current knowledge on the cognitive processes in the effect of eWOM on post-purchase responses, as well as explaining the mixed findings in the literature. Coherent with the ECT, the findings confirm the effects of expectation and confirmation on post-purchase responses where confirmation plays a more dominant role in the explanatory power. Specifically, eWOM valence has a positive effect on expectation and a negative effect on confirmation, as the effect of expectation is offset against that of confirmation, eWOM valence has no effect on post-purchase responses. EWOM volume has no significant effect on either expectation or confirmation, thus no effect on post-purchase responses. We argue that the effect of eWOM volume was not observed because the number of reviews indicates the level of perceived risk during the purchase, which was not the case in the free cost try-it-out online experimental scenario (Babić Rosario et al., 2016). Therefore, a more realistic purchase scenario is expected in future studies. EWOM quality has a positive effect on confirmation, which subsequently results in a positive effect on post-purchase evaluation. In sum, eWOM valence and quality play an important role in pre-purchase and post-purchase responses respectively, where confirmation is the dominant mediator. As the most immediate indicator of consumers’ post-purchase affective, attitudinal and behavioral consequences (Oliver, 1980), the effect of confirmation stayed unexplored in previous marketing research, which should attract more attention from academics.
Third, most research on eWOM effects have treated eWOM as an input, while the dynamic nature of eWOM as an output generating from post-purchase responses are largely
