The impact of product information on
preferences in online shopping environments:
the moderating role of consumer-generated
information
Author: Heijo ter Veldhuis
2
The impact of product information on
preferences in online shopping environments:
the moderating role of consumer-generated
information
University of Groningen
Faculty of Economics and Business MSc Marketing Management
Author: Heijo ter Veldhuis
Date: June 22, 2015
Adress: Schultestraat 64
9406 NG Assen
Phone number: +31(0) 623419656
Email address: h.w.ter.veldhuis@student.rug.nl Student number: 2584549
3
Management summary
Recent research underlines the importance of optimizing product information in online shopping environments. By providing customers with a lot of information, online retailers strive to give the customer the right product information on their e-commerce sites. Commonly, firm-generated information (i.e. product features and price), as well as consumer-generated information (i.e. ratings and reviews) are used to directly influence purchase behaviours. However, interaction effects between these two types are often not considered. Given the importance of these two types of information on the decision process of customers, online retailers should present their product information in such a way that it is most beneficial to the customer.
For this reason, this research focuses on further understanding the effects of consumer- and firm-generated information elements. To understand how these types of product information influence customer preferences, it is studied how firm-generated and consumer-generated information elements can be used to optimize the display of information on web pages.
Results of this study show the importance of review valence in aggregated ratings. Online retailers should be aware of the importance of this attribute and optimize their e-commerce environments to optimally display customer reviews to facilitate customer decision-making. In addition, the importance of price was shown to be moderated by the number of reviews. When the number of reviews is high, the price attribute becomes less important. Further research showed that not only the number of reviews, but the combined effect of review valence and number of reviews influence the importance of price. Online retailers should be aware of these effects and experiment with pricing strategies, conditional on customer reviews.
4
Abstract
In recent years, there has been an increasing interest in the value of consumer-generated information in online shopping environments. The literature defines two types of information: firm-generated information (FGI) and consumer-generated information (CGI). However, little attention has been given to the interaction effects between CGI and FGI. To be able to understand how to optimize the presentation of product information in order to influence consumer preferences, it is essential to understand how these types of information influence each other. To study these interaction effects, a choice-based conjoint analysis was performed. Results show that review valence was by far the most important dimension. In addition, the results revealed interaction effects between the number of reviews and price. Further analysis showed that price is seen a most important when combined with a low number of reviews. These results bring implications for online retailers to optimize their information presentation in online shopping environments. Moreover, this research contributes to the current literature that studies consumer-generated information in online shopping environments. In addition, it introduces the diagnostic format as a new variable for information presentation. Finally, this study contributes to the literature by providing bases for future research regarding information presentation in online shopping environments.
5
Preface
This thesis marks the end of my master Marketing Management. After a lot of work and hours put into writing this report, it is time to finish my study and start my career.
First of all, I would like to thank my supervisor Dr. Hans Risselada for his guidance and feedback during this last semester. Furthermore, I would like to thank my thesis group for the feedback moments and discussions we had about our work. Finally, I would like to thank my girlfriend, friends and family, which supported me during these final years of studying.
Enjoy reading.
6
Table of contents
Management summary ... 3 Abstract ... 4 Preface ... 5 Table of contents ... 6 1. Introduction ... 7 2. Theoretical framework ... 10 2.1 Product preference ... 11 2.2 Firm-generated information ... 12 2.3 Consumer-generated information ... 15 3. Research design ... 21 3.1 Research method ... 21 3.2 Data collection ... 21 3.3 Study design ... 23 3.4 Plan of analysis ... 25 4. Results ... 28 4.1 Sample Characteristics ... 284.2 CBC Analysis: main effects ... 29
4.3 CBC Analysis: interaction effects ... 33
5. Discussion ... 40
5.1 Theoretical implications ... 40
5.2 Managerial implications ... 43
5.3 Limitations and directions for further research ... 44
References ... 46
Appendix A: Initial results ... 51
7
1. Introduction
With the emergence of internet retailing, a major communication and transaction medium has risen for retailers to use as a channel to market and sell their products on (Holzwarth, Janiszewski, and Neumann 2006; Olbrich and Holsing 2012). In recent years, the importance of optimizing product information in online shopping environments has become evident. 78% of luxury shoppers research products before purchasing online (Google 2013). Furthermore, 92% of consumers trust earned media (e.g. recommendations from friends and family) above all other forms of advertising (Nielsen 2012). Online retailers account for this by providing customers with a lot of information on their e-commerce sites. While it is common to display firm-generated information (FGI) such as product features and price, consumer-generated information (CGI) such as ratings and reviews, is often used as a complement to further influence purchase behaviour (Baek, Ahn, and Choi 2012; Li and Hitt 2008). From an academic perspective, there is an increasing interest in the value of consumer-generated information (Ransbotham, Kane, and Lurie 2012). Research shows that positive CGI generates higher product sales by enhancing customer’s quality expectations and attitudes towards the product (Tang, Fang, and Feng Wang 2014). Furthermore, customer reviews affect product sales, drive viewership and are perceived to be more helpful than those written by experts (Chevalier and Mayzlin 2006; Li et al. 2013; Ransbotham, Kane, and Lurie 2012).
8
Previous research has studied how firm-generated information (e.g. price, product description) influences consumer choice (DiRusso, Mudambi, and Schuff 2011; Keller and Staelin 1987). Providing the right product information is important for online retailers to influence consumer preference (Hsee et al. 2009). Since FGI and CGI are displayed together in online shopping environments to influence buying behaviour, studying how these information types interact with each other may provide useful insights. The elements of CGI have received a lot of attention in the marketing literature. Researchers have investigated the role of electronic word-of-mouth in the form of product reviews and ratings (Chevalier and Mayzlin 2006; Ransbotham, Kane, and Lurie 2012). Another study investigated the relative impact of consumer reviews and ratings, assessing the effect of CGI on product sales (Chevalier and Mayzlin 2006). However, little attention has been given to the interaction effects between CGI and FGI. Understanding the interactions between CGI and FGI can provide more information on the relationships between these types of information and help to optimize the presentation of product information in online shopping environments.
This research focuses on understanding the interaction effects between CGI and FGI elements. To understand how different types of product information influence customer preferences, it is studied how firm-generated and consumer-generated information elements can be used to optimize the display of information on web pages. Through studying interaction effects between these information elements, it is investigated how product information can be presented in such a way that it is most beneficial for the consumer. The aim of this research is to answer the following research question:
“To what extent does consumer-generated information affect the relationship
between firm-generated information and product preference?”
Data for this study is gathered by means of a Choice-Based Conjoint analysis. To research combinations of different CGI and FGI elements and their interaction effects, 301 respondents were presented with choice sets of simulated e-commerce environments that had different combinations of consumer-generated and firm-consumer-generated information. Consequently, the respondents were asked to indicate which product they preferred most. The results of these choice tasks were then analysed to investigate the interaction effects between CGI and FGI and their effects on product preference.
9
most important when combined with a high number of reviews. In addition, the results suggest that the interaction between review valence and the number of reviews may be an influential determinant of the importance of price in online shopping environments.
This paper adds to the literature by further exploring the role of CGI and FGI in online shopping environments. The importance of consumer-generated information is emphasized, showing review valence to be an important decision heuristic. In addition, the number of reviews was shown to significantly interact with price. Price is seen a most important when there is a low number of reviews. Furthermore, building on previous research, the results showed that review valence and number of reviews are most appropriately considered as a joint construct. Analysis showed that the importance of price may depend on review valence and the number of reviews as a joint construct. These findings extend the current literature and provide new and interesting directions that can be explored in future studies.
10
2. Theoretical framework
Previous research on product information in online shopping environments has primarily focussed on understanding the determinants of purchase intentions and online sales. Especially the influence of consumer-generated information is a topic that is broadly represented in the literature. The effects of consumer reviews and consumer ratings on purchase intentions and sales have been researched extensively (Baek, Ahn, and Choi 2012; Chevalier and Mayzlin 2006; Park, Lee, and Han 2007; Senecal and Nantel 2004; Sridhar and Srinivasan 2012). Likewise, evidence suggests that firm-generated information such as price and product descriptions influence consumer choice (DiRusso, Mudambi, and Schuff 2011; Keller and Staelin 1987). Whereas the direct effects of consumer- and firm-generated information on consumer preferences, purchase intentions and sales have been broadly researched, the interaction effects between the two have received little attention.
This research focuses on the interaction effects between consumer-generated information and firm-generated information in online shopping environments. More specifically, it investigates the moderating effect of consumer-generated information on the relationship between firm-generated information and product preference. This is relevant to study due to the way that users consume information in online shopping environments. Research on information overload shows that increases in the amount of information lead to greater information processing, which leads to a decline in decision quality (Lurie 2004).
11
Figure 1: Conceptual framework
2.1 Product preference
In order to assess the importance of FGI and how its attributes are moderated by CGI, the impact of those variables on product preference is measured. In this study, product preference is defined as: ‘the preference of one product over another, as indicated by the consumer’. Consumers use information that is available online to update their beliefs about that product and their perception of how much they would enjoy using it (Bertini, Ofek, and Ariely 2009; Branco, Sun, and Villas-Boas 2012). Based on the information they gathered, they make the decision on whether or not to purchase the product. The preferred option is chosen based on the expected utility for that product (Archak, Ghose, and Ipeirotis 2011). Therefore, online retailers should strive to optimize the information they provide on the product in order to influence consumer preference (Hsee et al. 2009).
12
literature by studying the effects of firm-generated information on product preference and how these relationships are moderated by consumer-generation information. Furthermore, understanding how the importance of firm-generated information such as price and presentation format is influenced by consumer-generated information, can yield important implications for online retailers in managing their shopping environments. For example, if under certain conditions, price becomes less important, different pricing strategies can be used conditional on the information that is presented. Thanks to the flexibility and adaptability of e-commerce environments, different types of information presentation can be managed (Manvi and Venkataram 2005). Therefore, studying the relationships between information elements may result in new ways to make the presentation of information in e-commerce environments more adaptable. This would provide ways to arrange the display of firm-generated information, conditional on changes in consumer-generated information. This study strives to find several of this optimal combinations as input for online retailers to understand consumer preferences for information and to be able to respond to information needs of consumers.
2.2 Firm-generated information
One set of information that is important to the customer in online shopping environments is firm-generated information. This type of information is, as opposed to consumer-firm-generated information, provided by the firm and consists for example of price- and product description. Firm-generated information is defined as: ‘all the product information on an e-commerce website that is placed there by the retailer itself’. Providing in-depth product information is important and even expected by customers as a source to base their buying decision on (Eisingerich and Kretschmer 2008). Product information is an important decision variable for potential customers and is used by firms to influence purchase behaviour and product evaluations (Chang and Wildt 1994).
13
Price
The first aspect of firm-generated information is price. When buying online, price is an important factor in the customer decision-making process. Price has been researched to have a negative effect on perceived value and willingness to buy (Dodds, Monroe, and Grewal 1991). Therefore, it is assumed that price will have a negative effect on product preference. The following hypothesis is tested:
H1: The price of a product is negatively related to product preference.
Third-party rating
Another form of FGI through which consumers can inform themselves, is by reading third-party product reviews. These reviews are similar to online consumer reviews, only they are written by third parties. Third-party product reviews are usually based on lab testing or expert evaluations (Chen and Xie 2005). They are often used in industries such as consumer electronics, financial services and entertainment (Chen, Liu, and Zhang 2011). Examples of such sources are Tweakers.net (a review website for consumer technology products) and Rotten Tomatoes, which reviews movies. Third-party product reviews are usually accompanied with an overall rating for the product or service. This aggregate rating is defined in this study as third-party rating (TRR).
While consumer reviews are based on personal experiences, TRRs are more focussed on product attribute information such as performance, features and reliability. As a result, these ratings are likely to be correlated with the performance of a product or service on these attributes (Chen and Xie 2008). Helpful reviews should have a good perceived source credibility. This refers to the perceived credibility of the authors’ review information by the customer (Jain and Posavac 2001; Li et al. 2013). TRRs which are perceived as trustworthy can consequently be used as a trustmark. Trustmarks have been researched to have a positive effect on product evaluations and purchase intentions (Aiken, Shin, and Pascal 2014). Therefore, the influence of TRRs on product preference is assumed positive. This is reflected in the following hypothesis:
H2: The valence of a third-party rating is positively related to product preference.
Information presentation format
14
2005). On the other hand, research on information overload suggests that too much information can work counterproductive and cause great uncertainty for consumers (Aljukhadar, Senecal, and Daoust 2012; Markus and Schwartz 2010). Therefore, online retailers should carefully consider how to present product information (Hars et al. 2000). The format in which information is presented is defined as: ‘information presentation format’ (IPV).
Blanco et al. (2010) investigated the effects of various information formats in online product presentations and they describe two modes of textual product information: schematic- and paragraph mode. The former describes product characteristics as listed in tables or charts, whereas the latter is a full textual description of the product. In this study, paragraph mode is described as: ‘textual information presentation format’. Research shows that a schematic display of textual information improves perceptions of information quality, which causes users to remember information that appears schematically more easily (Blanco, Sarasa, and Sanclemente 2010). Following this research, a schematic product description is expected to have a greater effect on product preference than a textual product description. This results in the following hypothesis:
H3a: If product information is presented in a schematic format, the preference for the
product is larger than with a textual information presentation format.
Another form of presenting textual product information is displaying benefits and drawbacks about the product. This form of product information is emerging on a number of large e-commerce websites and is for example used by online retailers such as Coolblue to present advantages and disadvantages of a product in a structured way. Building on this concept, this research proposes a new mode of textual product information: diagnostic product description. In the context of online product reviews, content diagnosticity is referred to as the extent to which a review contains alternative interpretations of a problem and possible solutions to it (Herr 1991). Information that is highly diagnostic allows consumers to solve their problem more effectively and helps them to differentiate between the benefits and concerns about a certain product (Li et al. 2013). Moreover, information that is highly diagnostic, may help consumers in a specific decision-making task (Jiménez and Mendoza 2013). Following this reasoning, it is expected that a diagnostic product description compared to a schematic product description has a greater effect on product preference since it makes the information more relevant to the decision-making process. This results in the following hypothesis:
H3b: If product information is presented in a diagnostic format, the preference for the
product is larger than with a (1) textual- and (2) schematic information presentation
15
2.3 Consumer-generated information
One of the most important elements today in designing online shopping sites, is the use of consumer-generated information (CGI). Consumer-consumer-generated information consists of content that is contributed by consumers or users on an e-commerce platform. For example, CGI can consist of product ratings and product reviews. Consumer preferences are found to be influenced by online information such as reviews and ratings (Senecal and Nantel 2004). In this study, the effects of CGI on product preference are examined. Furthermore, it is investigated how the main effects of firm-generated information on product preference are affected by consumer-generated information. This study uses the following elements to account for consumer-generated information in online shopping environments: review valence, number of reviews and review variance. These three elements are widely used in research as a set of characteristics to capture consumer-generated information (Chintagunta, Gopinath, and Venkataraman 2010; Maeyer 2012). This section will discuss the hypothesized effects of these CGI elements and their interaction effects with FGI in more detail.
Review valence
Online product ratings are a key form of consumer-generated information. This type of CGI is widely used in online shopping environments to facilitate product evaluations. A product rating is a quantitative evaluation of a product and is usually displayed next to the product on a web page in the form of an aggregated rating (Sridhar and Srinivasan 2012). The display of this type of information is used to show a consensus by online reviewers’ evaluation of that product. An example of a product rating is the star rating, which can serve as a cue for the review content (Poston and Speier 2005). The average rating value of the reviews can be described as: ‘an aggregated value representing the valence of the reviews’ (Park, Gu, and Lee 2012). Research has shown that review valence (i.e. the average rating of the product) is positively associated with purchase intentions (Flanagin et al. 2014). The same research by Flanagin et al. (2014) shows that product ratings are critical in the evaluation of product information credibility in online shopping environments. Therefore, it is expected that an increase in review valence strengthens the quality perceptions for that product, which positively influences the preference for that product. This results in the following hypothesis:
H4: Review valence is positively related to product preference.
Number of reviews
16
is an indication of a higher number of people who bought the product. Research on data from amazon.com suggests that a higher number of reviews over time is associated with greater relative sales over time (Chevalier and Mayzlin 2006). Moreover, it has been researched that consumers’ purchase intention increases together with the number of reviews (Park, Lee, and Han 2007). The reasoning behind this is that a high quantity of reviews is an indicator of product popularity, which consequently increases purchase intention. This finding is supported by other studies, in which the results suggest that displaying the number of reviews next to a product or service, may enhance consumer perceptions of the product and increase the purchase intentions (Chen, Liu, and Zhang 2011; Zhu and Zhang 2010). Following this reasoning, it is assumed that a higher number of reviews may give customers a more positive perception of the product and by this means positively influences their preference. Following this reasoning, the following hypothesis is tested:
H5: The number of reviews is positively related to product preference.
Review variance
A number of studies have investigated at the impact of variance of product reviews. This is usually described as the variance of product ratings and is used, for example on amazon.com, to display the distribution in valence of product reviews (Maeyer 2012). The variance in product reviews typically shows a J-shaped distribution with some negatively-valenced reviews, a few average ratings and mostly positively-valenced reviews (Hu, Zhang, and Pavlou 2009). Research has shown that high variance in the reviews may simply reflect different preferences between segments (Sun 2012). On the contrary, recent research suggests that high review variance may lead to a higher perceived risk and may have a negative effect on sales (Maeyer 2012; Ye, Law, and Gu 2009). Following this reasoning, it is assumed that a high review variance shows a lack of consensus in the reviews which leads to a less significant effect of the review on consumer preference. In addition, research demonstrates that one important role of CGI in the decision process, is to decide whether a product is in the consideration set (Jang, Prasad, and Ratchford 2012). Therefore, a high review variance may cause the consumer to perceive the review as inconclusive, causing them to not even include the product in their consideration set. This may negatively influence the product preference. The proposed relationship is reflected in the following hypothesis:
H6: The variance of the review is negatively related to product preference.
Interaction effects with price
17
While CGI, such as the valence of the review, may directly influence product perceptions, it may also change the importance of price on the preference for that product. In the context of new product innovation, consumer-generated information has been shown to positively influence pay (Parry and Kawakami 2015). Furthermore, is has been studied that a consumers’ willingness-to-pay is influenced by the extent to which a decision is associated with risk (Yang, Vosgerau, and Loewenstein 2013).
An important antecedent of purchase decisions is perceived value. Prior studies suggest that purchase decisions are not necessarily affected by the objective price, but rather by the price perceptions of the customer (Danziger, Hadar, and Morwitz 2014; Zeithaml 1988). Perceived value is researched to be a construct that defines the relationship between price and purchase intention (Dodds, Monroe, and Grewal 1991; Zeithaml 1988). The trade-off between perceived price and perceived quality leads to perceived value (Chang and Wildt 1994). Prior research used review valence as an indicator of perceived value, reflecting the utility derived from the purchase in terms of quality and price. Hence, a positive review valence may reflect a high perceived quality by other customers and lessen the negative effect of price.
Following this reasoning, when confronting the customer with a more positive review valence, this may enhance their perceived value of the product, decreasing the negative effect of price on product preference. Therefore, the following hypothesis is proposed:
H7: The price of a product has a weaker effect on product preference when displayed
with a high review valence compared to a low review valence.
Having a higher number of reviews make online reviews to be perceived as more trustworthy (Zhu and Zhang 2010). The reasoning behind this is that a product that is more often reviewed, better reflects the true quality of the product. A higher number of product ratings can decrease the perceived risk (Flanagin et al. 2014). Furthermore, consumers’ willingness-to-pay may be influenced by perceived risk (Yang, Vosgerau, and Loewenstein 2013). Therefore, a higher product rating may reduce the perceived risk and increase the willingness-to-pay of the customer. Hereby decreasing the negative effect of price on product preference. Therefore, the following hypothesis is proposed:
H8: The price of a product has a weaker effect on product preference when displayed with
a high number of reviews compared to a low number of reviews.
18
review variance may lead to a lower willingness-to-pay, moderating the relationship between price and product preference. This leads to the following hypothesis:
H9: The price of a product has a weaker effect on product preference when displayed
with low variance reviews compared to high variance reviews.
Interaction effects with TRR
Review valence is an important factor in consumer decision-making. It has been studied that reviews written by consumers are perceived as more helpful than those written by experts (Li et al. 2013). Since review valence is consumer-generated and TRRs are written by third-party sources or experts, it is expected that consumers will most likely attach more value to review valence compared to third-party ratings. When consumers attach more value to consumer ratings, review valence is likely to be used as the dominant decision heuristic.
Research confirms the importance of review valence in the evaluation of information credibility (Flanagin et al. 2014). Furthermore, reviews are used as an important decision heuristic (Jang, Prasad, and Ratchford 2012). This research hypothesizes (H4) that a high review valence increases the quality expectations of a consumer for this product. Therefore, if the review valence is higher, the consumer will have less motivation to confirm product quality. Hence, making TRRs less necessary as an heuristic to base purchase decisions on. Following this reasoning, it is hypothesized that when consumers are confronted with a positive review valence, this may enhance their perceived value of the product, weakening the effect of a third-party rating. The following hypothesis is proposed:
H10: The valence of a third-party rating has a weaker effect on product preference
when displayed with a high review valence compared to a low review valence.
Although third-party ratings can be influential in consumer decision-making (Chen, Liu, and Zhang 2011), more determinants of consumer choice can be specified. As in the previous paragraphs, consumers may attach more value to consumer ratings compared to TRRs. As Zhu and Zhang (2010) explain in their paper, the larger the number of users that agree on the quality of a product, the more likely it is that the review represents the true quality of that product. When consumers are convinced of the true quality of the product, this may make expert reviews a less important attribute. Following this reasoning, when there is a high number of product reviews, this may increase the perceived product quality, weakening the effect of a third-party rating. This results in the following hypothesis:
H11: The valence of a third-party rating has a weaker effect on product preference
19
Research indicates that product evaluation is more positively influenced by low-variance product reviews compared to high-variance product reviews (Park and Park 2013). In addition, customer reviews are viewed as evaluations based on personal experience and are perceived as more helpful compared to expert-written reviews (Li et al. 2013). Therefore, it is expected that a low-variance review is perceived as a good representation of real life product quality and therefore weakens the impact of expert-written reviews, resulting in the following hypothesis:
H12: The valence of a third-party rating has a stronger effect on product preference
when displayed with low variance reviews compared to high variance reviews.
Interaction effects with IPV
Consumers that visit online shopping environments rely on product information to make purchase decisions. Earlier in this paper, it is hypothesized that a diagnostic or schematic presentation of information that is generated by the firm, is likely to be most influential on product preference. This positive effect may however be moderated by the valence of aggregated reviews. Research shows that product rating information is critical in evaluating the credibility of commercial information about the product (Flanagin et al. 2014). Subsequently, it is expected that the positive effect of firm-generated information about the product may be moderated by review valence in the sense that a positive review valence weakens the positive effect of information presentation. This results in the following hypothesis:
H13: The presentation format of product information has a weaker effect on product
preference when displayed with a high review valence compared to a low review
valence.
The number of product ratings have been researched to be a heuristic cue for the quality of consumer-generated information (Tang, Fang, and Feng Wang 2014). Consumers may use the number of product ratings as a mental shortcut to judge the credibility online as a way to deal with evaluating all available information (Flanagin et al. 2014). Following this line of thought, a high number of reviews may strengthen the importance of consumer-generated information (e.g. the valence of the review) and weaken the need for firm-generated information. Hence, the following hypothesis is researched:
H14: The presentation format of product information has a weaker effect on product
preference when displayed with a high number of reviews compared to a low number
20
A low variance in an information source, can increase its impact (Hu and Li 2011). In addition, CGI may be perceived as more trustworthy (Chen and Xie 2008). From this perspective, a low variance in the reviews may cause the consumer to view the consumer-generated information as more credible, because of consensus in the rating and a decreasing influence of firm-generated information. The following hypothesis is proposed:
H15: The presentation format of product information has a stronger effect on product
preference when displayed with low variance reviews compared to high variance
21
3. Research design
This section discusses the design of the research. First, the choice for the research method will be further explained. After that, the procedure of data collection is elaborated on. Finally, the study design and plan of analysis will be discussed.
3.1 Research method
This research is conducted by means of a Choice-Based Conjoint analysis (CBC). Conjoint analysis captures the evaluation of products by respondents by considering the product attributes and levels jointly. The preferences are then decomposed by using statistical methods (Eggers and Sattler 2011). A common used method in conjoint analysis, is CBC. With this method, respondents are shown sets of alternatives, from which they repeatedly have to select their preferred alternative (Mccullougii 2002). This type of analysis has proven to be an effective method since making choices is an important part of daily life. For this reason, CBC is an ideal method to assess the influence of information elements on product preference in online shopping environments. The analysis is a realistic method that can simulate the choices that are made on e-commerce sites when buying a product.
Hence, in order to analyse the preferences for information attributes in online environments, the CBC method will be used. In this study, FGI and CGI are considered as a set of attributes which influences the preference for a product. Each element of FGI and CGI is considered as a separate attribute which influences product preference. Consequently, the preference for a product is considered to be based on the utilities of attribute levels that are linked with that option.
3.2 Data collection
An online survey was created in order to distribute the research among respondents. The survey was created in Preference Lab and included the choice-sets from the CBC and several follow-up questions to measure control variables. The survey was then distributed by means of a direct mail. Furthermore, the survey was spread by sharing messages on social media networks such as Facebook, Twitter and LinkedIn.
Conjoint attributes
22
consumers (Aljukhadar, Senecal, and Daoust 2012). Therefore, it is expected that participants carefully consider the different information attributes and then make their decision. The attributes that were included in the conjoint analysis are shown in table 1. The following paragraphs describe the attributes in more detail.
Attribute Attribute levels
Price € 419 € 499 € 579 Information presentation format Textual format Schematic format Diagnostic format Third-party rating 1 out of 5 stars3 out of 5 stars 5 out of 5 stars Review valence 1 out of 5 stars 3 out of 5 stars 5 out of 5 stars Number of reviews 2 review
21 reviews 118 reviews Variance of the review Low-variance
Moderate-variance High-variance Table 1: Attribute levels conjoint
The attribute ‘price’ was manipulated by defining three attribute-levels with different prices. Desk-research on the price range for laptops showed that the average price for basic laptops ranges from 400 to 600 euro (Tweakers 2015). Therefore, the attribute-levels for price were constructed to range from €419 to €579. This way, the difference between the lowest and highest price was high enough to test the hypotheses, but not so large that it caused dominance in the choice sets. A pre-test (n=20), in which respondents were asked to indicate their maximum and minimum price, confirmed this assumption.
23
retailers (e.g. Coolblue, Wehkamp and Bol.com) in the Netherlands. From their websites, various laptops were inspected from which the number of reviews was derived. Building on this, the attribute-levels for number of reviews were stated to be 2, 21 or 118 reviews. Finally, the variance of the review was manipulated by displaying low-, moderate- or high variance in the review distribution.
Control variables
Several control variables are included in this study. In order to get a better understanding of the sample that is studied, the respondents were asked to answer several questions regarding their demographic characteristics. First, the respondents had to fill in their gender and age. Second, the educational level of the respondents was asked. Finally, the current employment status of respondents was evaluated.
3.3 Study design
The conjoint analysis that was used for this study, included 16 choice sets consisting of 3 alternatives each. Therefore, the choices of all 301 respondents represent 4816 choices. This high number of choice sets was used because this research is aimed at assessing interaction effects between attributes. To be able to measure these effects, a high number of respondents and choice sets was needed in order to find significant effects. Furthermore, the choice design was of high importance. An optimal choice design is balanced, orthogonal and minimizes overlap (Charzan and Orme 2000; Sawtooth Software 2013). However, recent studies in design efficiency for CBC, suggests that allowing some degree of overlap may improve the precision of interactions (Sawtooth Software 2013). The same article states that the random design method is considered most efficient in estimating interaction effects. Due to the goal of this research, assessing interaction effects between FGI and CGI, the random design method was used for the conjoint analysis.
Since the study contains of 6 attribute with 3 levels each, a total number of 729 (36) profiles could be created. Since this is a lot for respondents to process, a fractional factorial design was chosen. Hence, a fraction of the full number of alternatives was shown to respondents. The design of the choice-sets that was used in the surveys, was provided by the software of Preference Lab. The program provided the optimal design containing the choice tasks. Preference Lab accounted for the designs to be balanced, which means that each attribute is shown approximately the same number of times (Eggers and Sattler 2011). Moreover, the choice designs were created to be orthogonal. Hence, attribute-levels are measured independently from all other effects (Sawtooth Software 2013).
24
levels and displayed them within the alternative. The same was done for each of the attributes. Hence, each of the alternatives in the choice set consisted of 6 layers: one attribute-level from each of the six attributes. An example of such a choice set is displayed in figure 2.
Figure 2: Choice sets conjoint analysis
Procedure
25
3.4 Plan of analysis
To assess the results as described in the previous sections, the effects of the attributes on product preference is investigated. This is done by estimating the utility of each product level, which equals the sum of the utilities from each attribute level. Accordingly, the formula that determines the utility (U) of a product is:
U = ß
1PLOW + ß
2PMOD + ß
3PHIGH + ß
4TF + ß
5SF + ß
6DF + ß
7TRRLOW +
ß
8TRRMOD + ß
9TRRHIGH + ß
10RVLOW + ß
11RVMOD + ß
12RVHIGH +
ß
13NUMLOW + ß
14NUMMOD + ß
15NUMHIGH + ß
16VARLOW + ß
17VARMOD
+ ß
18VARHIGH
Where, U = Utility
26
In the analysis, three preference functions will be considered for each attribute. First, the part-worth model will be estimated. This function model estimates each attribute-level individually and therefore has the largest amount of parameters. The other preference functions are the linear model and the ideal-point model. Using one of these functions can make the model less complex, since these consider a lower number of parameters. To assess whether some variables can be treated as linear or ideal-point models, the part-worths for each attribute in the aggregate model will be plotted.
Model fit
The model fit is assessed to compare two different models to each other. This is done in order to analyse whether differences between the models, lead to significant improvements in model fit. To assess the goodness of fit of the model, an important information criterion is the AIC, which is based on the L2 values. The AIC take the parsimoniousness of the model into account by considering the number of parameters (Vermunt and Magidson 2005). Therefore, AIC is a well-suited measure for assessing model fit. When comparing AIC measures between models, a lower value represents a better model fit (Vermunt and Magidson 2015).
Validation
In order to see the prediction accuracy of the models, the internal and external validity can be assessed (Green, Carroll, and Goldberg 1981). The internal validity is assessed by analysing the predicted cases in Latent Gold, the statistical software in which the preferences are estimated. For measuring external validity, these values should be compared against a holdout set. A limitation of this research is that this value could not be compared to a holdout task. A holdout set was not included in the survey and can therefore not be compared to the hit rate of the aggregate model.
Relative importance
To assess the relative importance for each attribute, first the range is calculated. The range is based on the difference between the highest and lowest utility of an attribute and represents the maximum effect for that attribute (Vermunt and Magidson 2013). After this, the relative importance of each attribute is calculated by dividing its range with the total range (Vermunt and Magidson 2013).
Assessing interaction effects
27
of these attributes, the choice for a level of price is included conditional on the number of reviews. The new variables are then included in the model which results in different utilities for price, depending on the number of reviews.
28
4. Results
In this section, the results of the conjoint analysis are presented. First, the characteristics of the sample are discussed. Second, the main effects of the CBC analysis are shown. Finally, interaction effects between the attributes are examined.
4.1 Sample Characteristics
In total, 301 respondents participated in this research. In conjoint analysis, a sample size of at least 200 respondents is recommended for the group to represent the population (Hair and Black 2009). Hence, the number of participants in this research is sufficient for analysis. In table 2, characteristics of the respondents are summarized.
Characteristic
#
%
Gender Male Female 153 148 50.8% 49.2% Age < 16 16 – 25 26 – 35 36 – 45 45 – 55 56 – 65 65 + 0 210 49 16 23 3 0 0% 69.4% 16.6% 5.3% 7.6% 1% 0% EducationPrimary or secondary school MBO HBO WO 9 33 156 103 3% 11% 51.8% 34.2% Employment Employed Self-employed Currently unemployed Student Incapacitated 91 25 11 170 4 30.2% 8.3% 3.6% 56.5% 1.3% Table 2: Sample characteristics
29
years old representing 69.4% of the sample. Furthermore, the mean age of the overall sample is 27 and the median age is 23. This is in line with students representing 56.5% of the sample. Furthermore, 30.2% of respondents was employed, 8.3% was self-employed and the remaining respondents were either unemployed (3.6%) or incapacitated (1.3%). In addition, the sample consisted for the greater part of highly educated people, with 86% of respondents having a HBO or WO education. The remaining participants had an educational background in either MBO (11%) or primary/secondary school (3%).
4.2 CBC Analysis: main effects
This section discusses the aggregate model of the Choice-based Conjoint analysis. Insights in the relative importance of the attributes are provided. In addition, the main effects are analysed and compared to the hypothesized relationships.
Initial model
Initial estimation of the model showed that the attribute ‘review variance’ was highly insignificant (p=0.63, see appendix A). Omitting review variance from the model resulted in a new model. When comparing the initial model with the new model (without review variance), the new model showed a slightly higher adjusted R2, indicating that the model marginally better explains the variance in the model. Moreover, the model without review variance, showed a lower AIC value compared to the initial model (AIC: 7442 < 7445). Thus, model fit is improved when excluding review variance from the model. The results of the model comparison are summarized in table 3. In line with the findings, the attribute ‘review variance’ was excluded from further analyses. Accordingly, preferences are estimated from the remaining five attributes.
Model LL
Npar
Hit rate
R²adj
AIC
Initial model -3710.68 12 66.67% 0.296401 7445.37 Model without review variance -3711.15 10 66.59% 0.296691 7442.30 Table 3: Model comparison without Review VarianceModel selection
30
attribute ‘number of reviews’, does not indicate a linear relationship since the distance between 2 and 21 reviews is a lot larger than the distance between 21 and 118 reviews (see figure 3). Therefore, the number of reviews is considered as a part-worth function in the model.
Figure 3: Part-worth's of the attributes
As can be seen from figure 3, the attribute ‘pricing’ seems to show a linear relationship. To assess whether this attribute should be included in the model as linear, a new model is made and compared against the aggregate model. Table 4 shows the model comparison between the aggregate model and the model with pricing estimated as linear. Analysis of the results shows that the model with pricing included as linear, is different in fit compared the aggregate model. The adjusted r-square and hit rate show slight increases in the linear model compared to the aggregate model. Moreover, the linear model shows an increase in AIC compared to the aggregate model (7454.95 > 7442.30). Therefore, it can be concluded that including pricing in the model as a linear function, will not lead to a same or better model fit. Hence, all attributes are considered as part-worth functions in the final model.
Model LL
Npar
Hit rate
R²adj
AIC
Aggregate model -3711.15 10 66.59% 0.296691 7442.30 Pricing linear -3718.48 9 66.42% 0.295495 7454.95 Table 4: Model comparison aggregate versus linear models
Aggregate model
After excluding review variance from the model, the part-worth’s were estimated for the remaining attributes. Table 5 provides an overview of all the attributes with the part-worth values for each attribute-level. Each attribute in the aggregate model was showed to be significant. It can be concluded
-1,5 -1 -0,5 0 0,5 1 1,5 1 s ta r 3 s ta rs 5 s ta rs 41 9 49 9 57 9 textu al sch em at ic d iagn o stic 1 s ta r 3 s ta rs 5 s ta rs 2 re vie w s 21 re vi ew s 11 8 rev ie w s
31
that review valence is the most important attribute with a relative importance of 44.90%. Price is the second most important attribute with a relative importance of 20.75%. The number of reviews are also an important attribute with a relative importance of 15.64%. Finally, third-part review rating has a relative importance of 13.03% and information presentation format is perceived as least important attribute with a share of 5.86%.
Attribute
Part-worth
P-value
Range
Importance
Price 419 euro 499 euro 579 euro 0.5147 0.0932 -0.6079 0.000 1.1234 20.75% Third-party rating 1 star 3 stars 5 stars -0.3572 0.0088 0.3484 0.000 0.7053 13.03%
Information Presentation Format Textual Schematic Diagnostic -0.1886 0.0699 0.1187 0.000 0.3075 5.68% Review Valence 1 star 3 stars 5 stars -1.3346 0.2369 1.0977 0.000 2.4313 44.90% Number of reviews 2 reviews 21 reviews 118 reviews -0.4923 0.1373 0.355 0.000 0.8471 15.64%
Table 5: Aggregate model
Hypothesis testing
To assess the outcomes on the main effects, the effects of the attribute-levels in the aggregate model were analysed. Figure 3 and table 5 provide an overview on the estimated part-worth’s for each attribute-level. This shows that, relative to the mean, a high price has a negative effect (€579, β -0.61) on product preference, whereas an average price (€499, β 0.09) and high price (€499, β 0.51) have a positive effect. This shows that a high price compared to an average and low price, has a negative effect on product preference. Thus, supporting hypothesis 1. Furthermore, the results show that a high third-party rating (5 stars, β 0.34) has a positive effect, whereas a low rating (1 star β -0.35), has a negative effect on product preference. Hereby providing support for H2.
32
This supports H3a and H3b. A high review valence had a significantly positive effect on product preference (5 stars, β 1.10). Likewise, moderately-valenced reviews had a positive utility (3 stars, β 0.24), followed by a negative effect of negatively-valenced reviews (1 star, β -1.3). Hence, H4, is supported. The utility for the number of reviews was found to be positive for the highest number of reviews (118 reviews, β 0.36), with a marginally positive effect of a moderate number of reviews (21 reviews, β 0.14) and a negative effect of little reviews (2 reviews, β -0.49). Hereby supporting H5. Finally, the variance of the reviews was found to be insignificant in predicting product preference. A summary of the hypotheses on the main effects and their support in the model is depicted in table 6. These outcomes will be further discussed in the discussion section.
Hypotheses main effects
Overall support
H1: The price of a product is negatively related to product preference. Supported
H2: The valence of a third-party rating is positively related to product
preference.
Supported H3a: If product information is presented in a schematic format, the preference
for the product is larger than with a textual information presentation format.
Supported H3b: If product information is presented in a diagnostic format, the preference
for the product is larger than with a (1) textual- and (2) schematic information presentation format.
Supported
H4: Review valence is positively related to product preference. Supported
H5: The number of reviews is positively related to product preference. Supported
H6: The variance of the review is negatively related to product preference. Not supported
Table 6: Support for hypotheses main effects
Validity
33
4.3 CBC Analysis: interaction effects
In this section, the results with regard to attribute interactions will be discussed. The proposed hypotheses with regard to attribute-moderators will be tested and further analysed.
Model selection
To estimate the preferences of respondents while accounting for interaction effects between the attributes, the interactions are included in the model. Since review variance was tested as insignificant, this will not be taken into account. To assess each moderating effect, the variables were recoded as interactions and included in a model. Six new models were created that captured the interaction effects as discussed in the hypotheses. These models were then each compared to the model that did not include interactions in order to assess if adding interactions improves model fit. Table 6 provides an overview of these comparisons between the interaction models and the model without interactions. The last column summarizes for each interaction model, if an improvement was observed compared to the model that did not include interactions.
The greater part of the newly estimated models showed a small increase in their log-likelihood values when compared to the non-interaction model. In addition, there were no considerable differences in hit rate or adjusted R2 values. To assess the quality of each model relative to the non-interaction model, the AIC measure is used. From table 7 can be seen that the only model that shows a better AIC value compared to the non-interaction model is model 3. This included interactions between price and the number of reviews. The remaining models all show marginally higher AIC values compared to the aggregate model.
Model
LL
Hit rate
R²adj
AIC
Improved
model fit?
Model without interactions -3711.15 66.59% 0.2967 7442.30Model 2 (Price x valence) -3709.17 66.65% 0.2963 7446.34 No
Model 3 (Price x #reviews) -3706.75 66.57% 0.2968 7441.51 Yes Model 4 (TRR x valence) -3709.05 66.34% 0.2963 7446.09 No
Model 5 (TRR x #reviews) -3707.17 66.86% 0.2967 7442.34 No
Model 6 (Format x valence) -3710.84 66.45% 0.2960 7449.68 No
Model 7 (Format x #reviews) -3707.7 66.86% 0.2966 7443.40 No
Table 7: Model comparison interaction effects
Hypothesis testing
34
price in combination with negative-, moderate- or positively-valenced reviews. The following paragraphs will describe the results in more detail.
Interaction effects with price
The comparison of interaction the aggregate model showed that including interactions of price with review valence did not improve the model fit. Therefore, it can be concluded that review valence does not significantly interact with price to influence product preference. Hence, H7 is not supported. When interactions of price and number of reviews were included, the model did show an increase in model fit compared to the aggregate model (see table 6). To assess the moderating effects of number of reviews on pricing, the recoded variables (including interaction effects) were included in the model. As described in the research design section, the results were used to calculate the relative importance of pricing under different conditions of review valence. This provided an overview of the part-worth’s for pricing under different conditions for the number of reviews. Figure 4 provides an overview of the interaction effects between the number of reviews and pricing.
Figure 4: Moderating effects pricing x number of reviews
When interacting with different numbers of reviews, pricing shows a convincing difference in effects. From figure 3, it can be seen that when the number of reviews is high, the negative effects of a high price are weakened (-0.54 compared to -0.73 of low number of reviews). When put into terms of relative attribute importance, pricing shows to be a more important determinant of product preference when combined with a low number of reviews compared to a high number of reviews. Therefore, the results confirm H8 which states that price has a weaker effect on product preference when displayed with a high number of reviews compared to a low number of reviews.
419 499 579 2 reviews 0,6643 0,064 -0,7283 21 reviews 0,4834 0,1112 -0,5947 118 reviews 0,4265 0,1167 -0,5432 -0,8 -0,6 -0,4 -0,2 0 0,2 0,4 0,6 0,8 PART -W O R TH 15,00% 17,50% 20,00% 22,50% 25,00% Low #reviews Moderate #reviews High #reviews
Relative importance
35
Interaction effects with TRR
Since the models with interactions of TRR with number of reviews and review valence, produced no significant improvements (see table 7), no evidence was found to support H10 and H11. However, since there was only a slight increase in AIC, the outcomes of the interactions will be shortly discussed to show some intuitive results.
To gain more insights on the possible relationship between review valence and third-party rating, their interaction effects were assessed. From the results in figure 5 can be seen that for a low review valence, the effects of TRR are relatively close to each other. Whereas for moderately- and positively-valenced reviews, the positive effect of a good expert rating become greater and the negative effects of a bad expert rating become more negative. Horizontally, levels of TRR are shown and the columns show levels of review valence. The results in terms of relative importance show that the influence of TRR on product preference is the greatest when combined with a moderately-valenced review and the smallest when shown with a negatively-valenced review. However, these results should be carefully interpreted since no significant improvement was found in the model.
Figure 5: Moderating effects TRR x review valence
Interaction effects with IPV
Both the model that included interactions of review valence with information presentation format, as well as the model including interactions of number of reviews with information presentation format, produced no improvements in model fit. Hence, no evidence was found to support H13 and H14.
Interaction effects review valence and number of reviews
To assess the interaction effects between review valence and the number of reviews, first the interactions are included in a new model which was then compared to the model without interactions. From table 8 can be seen that adding valence and the number of reviews as an interaction in the model,
1 star (TRR) 3 stars (TRR) 5 stars (TRR)
36
improves model fit (AIC decreases from 7442.30 to 7337.14). In addition, hit rate and adjusted R2 measures improve for the newly estimated model compared to the model without interactions.
Model
LL
Hit rate R²adj
AIC
Improved
model fit?
Model without interactions -3711.15 66.59% 0.2967 7442.30Model 8 (Valence x #reviews) -3654.57 67.01% 0.3066 7337.14 Yes Table 8: Model comparison interaction-model with aggregate model
The output of this model was then used to display the relative importance of review valence dependent on the number of reviews. The results are shown in figure 6. From the figure can be seen that the valence of the review has the lowest attribute importance when the number of reviews is low. Moreover, review valence is perceived as most important when the number of reviews is high. Thus, confirming H16.
Figure 6: Moderating effects review valence x #reviews
Multiple-moderation effects
From the results above can be seen that review valence is a significantly better predictor of product preference when combined with a high number of reviews. Although the hypothesized relationships described review valence and the number of reviews as separate constructs, this study shows that the two are strongly related. Therefore, the hypothesized relationships of price, TRR and IPV may potentially not only be moderated by the separate effects of review valence and the number of reviews, but by the interaction between these two variables. This would result in a relationship where for example the importance of price is moderated by review valence, which is consequently moderated by the number of reviews. To test this assumption, some exploratory research was performed to give an indication whether this proposed relationship may be a fitting explanation of the relationships between FGI and CGI.
1 star 3 stars 5 stars
2 reviews -0,81 0,10 0,70 21 reviews -1,38 0,23 1,15 118 reviews -1,70 0,30 1,40 -2,40 -1,80 -1,20 -0,60 0,00 0,60 1,20 1,80 PART -W O R TH 25,00% 30,00% 35,00% 40,00% 45,00% 50,00% 55,00% Low #reviews Moderate #reviews High #reviews
Relative importance
37
First, the number of reviews and review valence were transformed into one variable that included their interaction effects. Then, the model fit was assessed. New models were created that included the valence x number-of-reviews interaction and its effect on an element of firm-generated information. This resulted in three different models that each included interactions between the new variable and price, TRR or IPV. Table 9 shows the comparison of the new models with the model without interactions and assesses the improvement in model fit. The results show that including valence and the number of reviews together as a moderator for price, significantly improves model fit compared to the model without interactions (AIC decreases from 7442.30 to 7431.84). Furthermore, hit rate and adjusted R2 measures improve.
Model
LL
Hit rate R²adj
AIC
Improved
model fit?
Model without interactions -3711.15 66.59% 0.2967 7442.30Model 9 (Valence and #reviews x Price) -3689.92 67.07% 0.2977 7431.84 Yes Model 10 (Valence and #reviews x TRR) -3695.94 66.88% 0.2965 7443.87 No
Model 11 (Valence and #reviews x IPV) -3702.90 67.15% 0.2952 7457.81 No
Table 9: Model comparison interaction-effect-models with aggregate model
For the model with the moderating effects of valence and number of reviews on price, the output was further analysed. Moreover, the relative importance of price was calculated under different conditions. Figure 7 shows the relative importance of price when shown with combinations of review variance and number of reviews. From the table can be seen that the relative importance of price differs not only under different conditions of review valence, but that this importance depends on the number of reviews. The figure shows that the relative importance of price is the highest when combined with a low-valenced review and a low number of reviews. In addition, the relative importance of price is the lowest when shown with a high-valenced review and a high number of reviews or a moderate valence and a low number of reviews.
38
Figure 7: Relative importance of price interacting with valence and #reviews
Summary of hypotheses
In the previous sections, the results of the choice-based conjoint analysis were presented. The results showed that the price of a product is influenced by the number of reviews. Price is a more important attribute when displayed with a low, compared to a high number of reviews. Furthermore, the results supported the hypothesis (H16) that review valence has a stronger effect on product preference when combined with a high, compared to a low number of reviews. This interaction between number of reviews and review valence was then further explored by taking the interaction effects into account when evaluating the importance of firm-generated information. In addition, results indicated that the importance of price not only depends on the number of reviews, but that it may be more significantly influenced by interactions between review valence and number of reviews. Table 10 provides an overview of the hypotheses and findings on the interaction effects that were discussed.
Hypotheses moderating effects
Overall support
H7: The price of a product has a weaker effect on product preference when
displayed with a high review valence compared to a low review valence.
Not supported
H8: The price of a product has a weaker effect on product preference when
displayed with a high number of reviews compared to a low number of reviews.
Supported H9: The price of a product has a weaker effect on product preference when
displayed with low variance reviews compared to high variance reviews.
Not supported
H10: The valence of a third-party rating has a weaker effect on product
preference when displayed with a high review valence compared to a low review valence.
Not supported
Low valence Moderate valence High valence
39
H11: The valence of a third-party rating has a weaker effect on product
preference when displayed with a high number of reviews compared to a low number of reviews.
Not supported
H12: The valence of a third-party rating has a stronger effect on product
preference when displayed with low variance reviews compared to high variance reviews.
Not supported
H13: The presentation format of product information has a weaker effect on
product preference when displayed with a high review valence compared to a low review valence.
Not supported
H14: The presentation format of product information has a weaker effect on
product preference when displayed with a high number of reviews compared to a low number of reviews.
Not supported
H15: The presentation format of product information has a stronger effect on
product preference when displayed with low variance reviews compared to high variance reviews.
Not supported
