RISK PERCEPTION

ORGANIZING ONLINE CONSUMER REVIEWS:

A CONJOINT ANALYSIS OF AGGREGATED PRESENTATION ATTRIBUTES AND THE MODERATING ROLE OF PRODUCT

RISK PERCEPTION

GIEL LINTHORST January, 2014

MSc Marketing Intelligence / Management University of Groningen

Faculty of Economics and Businesses - Department of Marketing

2

ORGANIZING ONLINE CONSUMER REVIEWS:

A CONJOINT ANALYSIS OF AGGREGATED PRESENTATION ATTRIBUTES AND THE MODERATING ROLE OF PRODUCT

RISK PERCEPTION

Author: Giel Linthorst

Sledemennerstraat 39-3 9718 BX Groningen Tel: +31 (0)630601382

Student number: 1787500

Supervisor: Dr. H. Risselada 2

Supervisor: MSc. L. de Vries

University: University of Groningen

Faculty of Economics and Businesses – Department of Marketing MSc Marketing Intelligence / Management

Research Theme: Online Consumer Reviews

3

MANAGEMENT SUMMARY

This research examines the relationship between online review presentation attributes and review page usefulness. This study gives valuable insights in which aggregated review attributes are the most important contributors to review page usefulness. Currently, a vast majority of review pages on websites are suboptimal. Implementing the right formats, showing the right information and providing the right options to customers on review pages increases the usefulness for customers. Increased usefulness can lead to an increase in sales since customers have better purchase support and guidance from the online review pages.

The usefulness of review pages is influenced by the way online consumer reviews are organized. The presentation format of online reviews, the provision of additional search options and presenting aggregated review information such as ratings, dispersion (variation in ratings) and summaries to customers have been included in this study.

232 Dutch respondents participated in an online choice-based conjoint analysis in order to measure preferences and relative importance of the presentation attributes. Preferences among respondents are estimated on both an aggregated and segmented level.

Results show that the presence of aggregated review information attributes (Rating, Dispersion and Summary) is the most important contributor to review page usefulness followed by presentation format and options for further research. A sophisticated presentation format whereby the most useful rated positive review is contrasted with the most useful rated negative review is more useful than less sophisticated review formats based on usefulness ratings or date. “See more reviews of similar products” is the most preferred option for further research.

No support for the moderating effect of product risk perception on the relationships between the presentation attributes and review page usefulness is found.

Key words: Online consumer reviews, rating, dispersion, summary, presentation format, review

page usefulness, review usefulness, product risk perception, choice-based conjoint

analysis.

4

ACKNOWLEDGMENTS

This master thesis marks the end of my Master Marketing Intelligence / Management program as well as my time as a student in Groningen. I will look back on the past years as a student with a feeling of great gratitude and satisfaction. Writing this thesis has been very instructive, but sometimes challenging as well. I would like to thank several people who have supported and assisted me in order to complete this master thesis.

First of all, I would like to thank my supervisor Dr. Hans Risselada for his guidance and support during the previous months. Furthermore, I would like to thank my fellow thesis group students for their constructive feedback, idea sharing and numerous coffee breaks in which we discussed our findings and challenges.

Also I would like to thank my respondents for their willingness to fill out my survey. Without their help, I would have been unable to gather the necessary data and subsequently finish this thesis.

A special thanks to my family, friends and my girlfriend for their support during the process. I would like to dedicate this master thesis and express my greatest gratitude to my parents Gerard and Anneke Linthorst who have supported me unconditionally over the past years.

Giel Linthorst

January, 2014

5

TABLE OF CONTENTS

MANAGEMENT SUMMARY ... 3

ACKNOWLEDGMENTS ... 4

1. INTRODUCTION ... 7

2. THEORETICAL FRAMEWORK ... 7

2.1 WOM and eWOM ... 10

2.2 Online consumer reviews (OCR) ... 10

2.2.1 Review page usefulness, review usefulness and usefulness ratings ... 11

2.2.2 Presentation format of OCR ... 12

2.2.3 Aggregation of OCR information ... 14

2.2.4 Options for further research ... 15

2.2.5 Product risk perception ... 16

2.3 Conceptual model ... 18

3. RESEARCH DESIGN ... 19

3.1 Methodology ... 19

3.1.1 Model specification ... 19

3.2 Stimuli ... 20

3.3 Design ... 21

3.3.1 Hold-out sample ... 23

3.3.2 Interaction effects ... 23

3.4 Product risk perception ... 23

6

3.5 Control variables ... 24

4. RESULTS ... 25

4.1 Descriptive statistics ... 25

4.2 Scale validation product risk perception ... 26

4.3 Conjoint analysis ... 28

4.3.1 Interaction effects ... 30

4.4 Hypotheses testing ... 31

4.5 Predictive validity ... 31

4.6 Segments ... 32

4.6.1 Interpretation of the segments ... 34

4.7 Overview of hypotheses ... 38

5. DISCUSSION ... 39

5.1 Implications ... 40

6. LIMITATIONS AND FUTURE RESEARCH ... 42

7. REFERENCES ... 43

8. APPENDICES ... 49

Appendix 1 : Online Survey (Qualtrics.com) ... 49

Appendix 2 : Descriptive statistics of the segments ... 61

7

1. INTRODUCTION

Due to the expanding reach of the internet, electonic word of mouth (eWOM) has become very important in online decision making. Word of mouth has proven to be of significant importance on product sales (Zhu & Zhang, 2010). Therefore, WOM has to be taken into account by firms (Chintagunta, Gopinath, & Venkataraman, 2010). Online consumer reviews (OCR) are considered to be one of the most important elements of eWOM (Schindler & Bickart, 2005). OCR is defined as a new piece of information presented from the perspective of consumers who have purchased and used the product (Park, Lee, & Han, 2007). This piece of information can serve as an important element in order to identify the product that best matches one’s personal preferences (Chen & Xie, 2008).

Online consumer reviews have become very accessible to consumers. Research has shown that 68% of the online shoppers is reading at least four reviews before making a purchase (Godes & Silva, 2012).

Consumers rely on user-generated content for several reasons; consumers consider user-generated content less intrusive than producer-generated content (Winer, 2009) and information retrieved from customers is also generally perceived as more credible (Bickart & Schindler, 2001). In 1995, amazon.com started to offer consumers an option to post their comments on products. In 2008, online consumer reviews on amazon.com were already exceeding 10 million (Chen & Xie, 2008). This increase in OCR volume goes hand in hand with the increase in popularity of OCR and the increased reliance on OCR (Zhu & Zhang, 2010). Numerous products on amazon.com have nowadays more than 1000 individual reviews. However, the high volumes of consumer reviews make it harder for customers to distinguish the best reviews from the mass and to understand the true underlying quality of a product based on the reviews (Ghose & Ipeirotis, 2007).

The increase in the amount of product reviews is therefore not necessarily an improvement. In fact, researchers collectively agree that more information does not always have a uniformly positive effect.

Meyer and Johnson (1989) argue that an increase in information can be harmful in the sense that

consumers make different choices than would have been made with broader processing power. Lurie

(2004) concludes similar findings in his research; an increase in the number of alternatives in a choice

set can lead to declining decision quality. Lee and Lee (2004) have studied the effect of information

overload on product sales in an online setting. The results imply that when consumers are exposed to

8 too much product information, they become information overloaded. This results in making poorer and less effective decisions and subsequently in a poorer product choice.

Organizing OCR can reduce the information overload and improve product choice decisions (Chen & Xie, 2008). The influence of several aspects of OCR has therefore been widely discussed in literature (Zhu &

Zhang, 2010). It is clear that individual review characteristics are determinants of review usefulness and product sales. Valence of OCR has proven to be of significant influence on sales (Chevalier & Mayzlin, 2006; Zhang, Ma, & Cartwright, 2013). Also volume (Liu, 2006) and dispersion (Clemons, Gao, & Hitt, 2006) of OCR influence sales. However, less is known about how customers process OCR (Purnawirawan et al., 2012) and how the presentation of OCR on review pages affects this.

Since consumers increasingly rely on OCR in their purchase decision, it is important for websites to organize their reviews in a way that maximizes customer usefulness. Usefulness of reviews is found to be very important; only when reviews are perceived as useful, the reviews lead to the formation of attitudes and (purchase) intentions (Purnawirawan, De Pelsmacker, & Dens, 2012). Websites should provide a consumer review page in which the usefulness of the consumer reviews is guaranteed. In this paper, review page usefulness refers to a webpage on a website that displays the online consumer reviews of a single product. Therefore, usefulness of reviews is examined on an aggregated level.

Websites present their reviews on review pages. These review pages present reviews by default in a certain way (for example: “most recent review first”). However, websites strongly differ in the way they present reviews. They provide different presentation attributes (formats, information metrics and options) to their customers. This paper investigates which of these presentation attributes are strong contributors to review page usefulness. The main research question is therefore as follows:

How should websites organize their reviews in order to maximize the usefulness of their review pages?

Moreover, this paper will investigate the moderating role of product risk perception on presentation

attributes. The more product risk is perceived, the more customers will search for additional product

information (Akaah & Korgaonkar, 1988) and will therefore rely more heavily on OCR since OCR is an

important piece of (additional) product information. This paper examines if there are differences in

9 valuation of presentation attributes among high and low product risk perceivers. High product risk perceivers are likely to look for risk-reducing activities (Dowling & Staelin, 1994). Therefore high product risk perceivers rely more heavily on OCR and benefit stronger from effective organization of OCR.

Moreover, high product risk perceivers will likely be willing to sacrifice more cognitive effort to examine the product reviews and therefore rely less on aggregated review information attributes.

An online choice based conjoint analysis has been conducted among 232 Dutch respondents in order to expose people’s preferences and investigate the contribution of each presentation attribute on review page usefulness. Results show that, in order to maximize review page usefulness, websites should provide a presentation format whereby the most useful rated positive review is contrasted with the most useful rated negative review, show Rating, Dispersion and a Summary and provide the option “See more reviews of similar products” for further research. No support for the moderating effect of product risk perception on the relationships between the presentation attributes and review page usefulness is found.

The results of this study have several contributions to the existing literature. First, this study examines the influence of OCR on an aggregated level. Thus, compared to previous studies, this study better reflects the actual processing situation when consulting OCR. Second, the presentation format of online reviews and the provision of options for further research have not yet been investigated in prior studies.

Third, the relative importance of presentation attributes is addressed, which gives valuable insights in the processing of OCR among customers. Fourth, the influence of product risk perception on OCR presentation preferences has not yet been investigated.

Practitioners can also benefit from this study. This study gives valuable insights in which aggregated review attributes are the most important contributors to review page usefulness. Currently, a vast majority of review pages on websites are suboptimal. Implementation of the right formats, showing the right information and providing the right options to customers on review pages increases the usefulness for customers. Increased usefulness can lead to an increase in sales since customers have better purchase support and guidance from the online review pages.

The remainder of this paper is structured in the following manner. First an overview of relevant existing literature is given. Secondly, an elaboration on the research design and data collection will be given.

Third, the results are presented and finally conclusions, implications and suggestions for further

research are discussed.

10

2. THEORETICAL FRAMEWORK

In this chapter, an overview of existing literature will be given to provide a deeper understanding of the main topic; organizing online consumer reviews. As a starting point the evolution of word of mouth (WOM) and electronic word of mouth (eWOM) and the development of OCR are being discussed. Based on the existing literature hypotheses are formed about the relation of presentation attributes on review page usefulness and the influence of product risk perception. The chapter will conclude with a conceptual model and an overview of all hypotheses.

2.1 WOM and eWOM

WOM involves informal communication among consumers about products and services (Liu, 2006).

Researchers have extensively investigated the relationship between WOM and product sales. Favorable WOM was found to increase the probability of purchase whereas exposure to negative comments decreased the probability of purchase (Arndt, 1967). WOM has a potential strong effect on product sales and customers rely on WOM as a tool for making their decisions (Herr, Kardes, & Kim, 1991). Due to the developments of the internet, WOM has increased significantly in importance. Not only has the internet increased the reach and accessibility of people, but it also caused an increase in the number of people exchanging or sharing their product information (Purnawirawan et al., 2012). With the use of the internet, consumers can easily publish their opinions, provide thoughts, feelings and viewpoints on products and services to the public at large (Schindler & Bickart, 2005). The face-to-face offline communication has transformed into eWOM. Recent scholars suggest that eWOM has become extremely important in the decision-making process of customers (Zhang et al., 2013).

WOM and eWOM have several similarities but they also differentiate on various aspects. The most important difference is that eWom is far more voluminous compared to WOM and the reach of eWOM goes far beyond the original WOM setting (Chatterjee, 2001; Dellarocas, 2003; Zhang et al., 2013).

2.2 Online consumer reviews (OCR)

OCR are considered one of the most important elements of eWOM (Schindler & Bickart, 2005). As

mentioned earlier, OCR plays a crucial role in making online product decisions (Zhu & Zhang, 2010).

11 Consumer-created information is considered more credible than seller-created information (Bickart &

Schindler, 2001) and several recent studies have taken the perspective of credibility of OCR. For example, Dellarocas (2003) has examined the relationship between online consumer feedback information on seller’s reputation. Moreover, OCR is found to be less intrusive than seller-created content (Winer, 2009).

Some scholars refer to both professional and customer reviews in their research about online reviews.

Extensive research has been conducted on the influence of professional reviews on customer choices and sales. For example, Basuroy, Chatterjee and Ravid (2003) found evidence that professional critics can influence and predict box office revenues of movies. However, the remainder of the current paper will focus on the reviews generated by consumers.

2.2.1 Review page usefulness, review usefulness and usefulness ratings

An important distinction has to be made between review page usefulness, review usefulness and usefulness ratings of individual reviews. In order to explain the differences among these concepts and their role in the hypotheses of this study, the concepts are listed below.

Review page usefulness

This paper investigates the relationship between presentation attributes and review page usefulness, the usefulness of a webpage on a website that displays the online consumer reviews of a single product.

Review page usefulness has received little or no attention from scholars. Elements of a review page, such as the balance and sequence of reviews (Purnawirawan et al., 2012), review volume (Liu, 2006) and review order (Purnawirawan et al., 2012) have been studied before, but a review page as a whole, where several presentation attributes are displayed, has not yet been studied. The influence of these presentation attributes on an aggregated level has thus not been investigated. Yet, these aggregated presentation attributes, for example: presentation format, options for further research and presence of rating, dispersion and summary, are widely used by practitioners.

Review usefulness

Although review usefulness is not investigated in this study, findings from prior research in this area

have been used in the formation of hypotheses. Review usefulness refers to the usefulness of a single

12 online consumer review. Review usefulness and the influence of review characteristics of individual reviews have received far more attention from researchers compared to review page usefulness.

Researchers have recently investigated the influence of online consumer review characteristics on sales and purchasing behavior. Recommendation (Senecal & Nantel, 2004), Ratings (Chevalier and Mayzlin, 2006), internet experience (Zhu & Zhang, 2010), review content (Archak, Ghose and Ipeirotis, 2011), review style (Ludwig et al. 2013) and brand strength (Ho-Dac, Carson and Moore, 2013) are all found to be influencing review usefulness.

Usefulness rating of individual reviews

Finally, the usefulness rating of individual reviews is being discussed. Contrary to review usefulness, this concept has been used in this study. Usefulness rating is a rating that (previous) customers have addressed to an individual review. Several websites allow readers to rate individual reviews on their usefulness. Most often, a “do you find this review useful? Yes / No” option is included or a similar

“thumbs up / thumbs down” option is provided to extract the usefulness rating of individual reviews from customers. As an example, amazon.com shows the usefulness rating of individual reviews in a “x out of x customers found this review useful” setting where other websites like bol.com and wehkamp.nl show the usefulness of individual reviews similarly (x times useful / x times not useful). Yahoo.com is showing usefulness ratings of individual reviews in a ratio (Hao, Li, & Zou, 2009).

The usefulness ratings of OCR make information retrieval and decision making more efficient for potential customers (Hao et al., 2009). Ghose and Ipeirotis (2007) have investigated the influence of usefulness ratings of individual reviews and they conclude that displaying the most useful individual reviews first significantly increases the review page usefulness for the users. Purnawirawan et al. (2012) confirm these findings by concluding that usefulness of individual reviews plays a crucial role. They argue that only when reviews are perceived as useful, the reviews lead to the formation of attitudes and (purchase) intentions.

2.2.2 Presentation format of OCR

As discussed, extensive research has been done in the field of OCR (Zhang et al., 2013; Zhu & Zhang, 2010). The focus of the majority of existing literature is on the content and characteristics of reviews.

Far less is known about the way websites organize their reviews and how presentation attributes affect

review page usefulness. One of the most apparent organizational attributes is the presentation format

13 of online reviews on webpages. Websites such as amazon.com, bol.com and hostelworld.com use various formats of presenting their reviews. Hostelworld.com is presenting their reviews standard based on date (newest first), bol.com (a Dutch online web shop) presents their reviews standard based on usefulness of reviews (usefulness rating of reviews) and amazon.com presents their reviews standard based on usefulness of the reviews in a combination with a star rating (“most useful rated positive review” is contrasted with “most useful rated negative review”). The presentation formats in this study differ from each other on their levels of sophistication whereas presentation based on date is the least sophisticated and the presentation format based on usefulness combined with a usefulness rating as the most sophisticated. The question then arises; do consumers perceive a more sophisticated presentation format as more useful?

In general, consumers cannot process all the information that is available for making a purchase decision (Duhan, Johnson, Wilcox, & Harrell, 1997). Duhan, Johnson, Wilcox and Harrell, (1997) argue that consumers use decision strategies in order to simplify a purchase. Recommendations (usefulness ratings) by consumers can be used for the purpose of reducing the amount of information that has to be processed to make a purchase decision. Rosen and Olshavsky (1987) have found that recommendations are used to reduce the number of alternatives and attributes that have to be considered. It is therefore expected that the presence of usefulness ratings of reviews will increase the usefulness of the review page. Review pages where reviews are ranked on usefulness rating are expected to have higher review page usefulness than webpages where reviews are ranked on date.

Recent research about the influence of OCR on sales has shown that the star rating (valence) of a review significantly affects sales (Chevalier & Mayzlin, 2006; East, Hammond, & Lomax, 2008; Liu, 2006; Zhang et al., 2013) and negative WOM is considered to be of greater impact than positive WOM (Chen, Wang,

& Xie, 2010). Extreme ratings are found to be of bigger influence than less extreme ratings (Clemons et

al., 2006). Chevalier and Mayzlin (2006) conclude that improvements in review ratings leads to increases

in product sales and the impact of a one-star rating is greater than the impact of a five-star rating. Since

the valence of reviews has proven to be one of the important drivers of purchase intention it can be

argued that the presence of valence in the presentation format of reviews has a positive effect on the

usefulness of a webpage. In this sense, a presentation format whereby the most useful rated positive

review is contrasted with the most useful rated negative review is expected to increase review page

usefulness. Therefore, a more sophisticated way of presenting reviews by including valence (most useful

14 rated positive vs. most useful rated negative) will increase review page usefulness. The following hypothesis is proposed:

H1: A more sophisticated format of presenting reviews to customers has a positive influence on review page usefulness.

2.2.3 Aggregation of OCR information

Besides the individual characteristics of OCR, aggregation of review information has been researched frequent and applied widely in practice. A discussed before, valence of OCR plays an important role in the decision process of consumers (Liu, 2006). Almost every website uses a star rating to express the valence of products on their website. The following hypothesis is therefore expected:

H2a: Showing a star rating of consumer product ratings as aggregated product information has a positive effect (compared to not showing a star rating) on review page usefulness.

Many websites and platforms like android’s app store show aggregated star ratings in combination with a distribution of the different star ratings. The question is if this combination between a star-rating and an indication of the dispersion of the ratings is perceived as more useful than only a star rating.

As the number of consumer reviews increases, the overall product rating converges to the true quality (Zhu & Zhang, 2010). This implies that ratings for popular products accurately reflect the true quality of the product and reliance on this rating is therefore less insecure. Purnawirawan et al. (2012) have studied the balance and sequence of online review sets on the usefulness of those sets. They find support for the findings that perceived usefulness is affected by the balance and sequence of review sets. Positive or negative balanced sets are considered to be more useful than neutral review sets.

Clemons et al. (2006) have also investigated the dispersion of reviews. They have found support for the

relation between dispersion and sales. In particular they find support for the influence of variance on

product growth; it is more favorable to have some customers who love your product than a huge

number of customers who predominantly like your product. These findings suggest that a highly

dispersed average rating of 3 stars differs from a less dispersed average rating of 3 stars.

15 From a psychological point of view, the credibility of websites can decrease if a clear positive set of reviews is displayed (Doh & Hwang, 2009). They also conclude that the credibility of websites can be harmed in the long term by clear positive review settings. The dispersion (variation) of reviews on an aggregated level gives insights in the consensus of reviews and could place a (positive) set of reviews of a review page into perspective. This suggests that showing the dispersion of reviews on an aggregated level could contribute to the credibility of websites and therefore contributes to the usefulness of the review page. Based on the findings in literature, the following hypothesis is proposed:

H2b: Showing a star rating and dispersion of consumer product ratings as aggregated product information has a positive effect (compared to only a star rating) on review page usefulness.

There are several websites who provide, next to star ratings, also a summary of the content of the reviews. For example hostelworld.com summarizes the scores of each reviewer on safety, location, staff etc. on an aggregated level. Bol.com summarized the most mentioned positive and negative emotions of their reviewers on an aggregated level at the top of the list of reviews. For example an aggregated book review summary contains pros: Inspiring (10), surprising (8) and cons; bad pictures (6), whereas the number in brackets refers to how often an emotion is mentioned in all of the reviews.

Some scholars argue that customers read review text rather than relying only on summary statistics (Archak, Ghose, & Ipeirotis, 2011; Chevalier & Mayzlin, 2006) and review depth has a significant effect on review usefulness (Mudambi & Schuff, 2010). A summary with the most frequent mentioned emotions of customers could be the link between aggregated information statistics and review text since the summary is derived from the actual text. Customers’ need for reading actual review text could therefore be reduced by a summary. A summary, derived from review texts, could reduce the cognitive effort and increase the usefulness of review pages. Therefore the following hypothesis is expected:

H2c: Showing the rating, dispersion and a summary of reviews as aggregated review information has a positive effect (compared to only a star rating and dispersion) on review page usefulness.

2.2.4 Options for further research

16 Another organizational presentation attribute on review pages is the provision of additional search options to customers. Websites like amazon.com and bol.com provide several options; “see more 1 2 3 4 5 star reviews”, “see more reviews of this reviewer” and “see more reviews of similar products”. The question is whether these options increase the overall usefulness of reviews.

As discussed before, valence has an important effect on sales (Chevalier & Mayzlin, 2006; East et al., 2008; Rosen & Olshavsky, 1987; Zhang et al., 2013). Extreme product ratings are considered of greater impact (Liu, 2006). Also negative eWOM has found to be a greater impact than positive eWOM (Park &

Lee, 2009). The option “see more 1 2 3 4 5 star reviews” gives customers direct access to extreme product ratings (1 and 5 star ratings) and negative reviews (1 and 2 star ratings). The desired product ratings (1, 2, 3, 4 or 5 star rating) can be shown with one mouse click. It is therefore expected that the provision of an option for further research that provides immediate access to these impactful reviews is valued as more useful. The following is hypothesized:

H3: The option “see more 1 2 3 4 5 star reviews” has a bigger positive influence on review page usefulness than the options “see more reviews of this reviewer” and “see more reviews of similar products”.

2.2.5 Product risk perception

Risk perception has been researched extensively in traditional decision making processes (Forsythe &

Shi, 2003). The perception of risk can be defined as the combination of the likelihood that something will go wrong and the perception of the seriousness of the consequences if it does (Kaplan, Szybillo &

Jacoby, 1974). Taylor (1974) has have investigated the role of risk on consumer behavior. He argued that the central problem of consumer behavior is choice and because the outcome is always unknown, the consumer is forced to deal with uncertainty, or risk.

Risk perception in online shopping has received more recent attention of scholars. Forsythe and Shi

(2003) distinguish four types of risk most relevant for online shopping; financial risk, product

performance risk, psychological risk and time/convenience risk. Financial risk is defined as the net loss of

money of a customer including potential credit card misuse risks (Horton, 1976). Product performance

risk is defined as the risk of making a poor product choice due to the shopper’s inability to accurately

17 judge the quality of the product (Forsythe & Shi, 2003). Psychological risk refers to disappointment, frustration and shame associated if one’s personal information is disclosed and also includes privacy risks. Finally, time/convenience risks are defined as the result of difficulties in navigating and submitting orders, finding appropriate websites or delays in receiving the product (Forsythe & Shi, 2003). Product performance risk is found to be the most important type of risk for online shoppers (Forsythe & Shi, 2003). Due to the size and scope of this research and the relative importance of product risk perception in online shopping, only the relation of product performance risk on review page usefulness is taken into account in the remainder of this paper.

Online shopping is perceived by consumers to be more risky compared to an offline purchase (Donthu &

Garcia, 1999). However, there is little known about the effects of this online product risk perception on consumer behavior (Forsythe & Shi, 2003). According to Taylor (1974), uncertainty involves two aspects:

uncertainty about the outcome and uncertainty about the consequences. Reducing the consequences through reducing the amount at stake can decrease uncertainty about the consequences. Uncertainty about the outcome can be reduced by handling and acquiring information. The second element, handling and acquiring information, is particularly interesting in this study. In fact, OCR have become an important element of product information (Zhu & Zhang, 2010) and OCR could very well reduce the uncertainty about the outcome by improving the acquiring and handling of product information.

Given that risk perception differs among individuals (Assael, 1987), it is likely that only those with a high product risk perception may seek formal and informal information sources to reduce their uncertainty (Akaah & Korgaonkar, 1988). Higher risk perception among customers increases the need for risk- reducing activities (Dowling & Staelin, 1994). Reading OCR can be a risk- reducing activity. High risk perceivers will experience higher barriers to shop online and therefore rely more heavily on informal information sources. OCR could be that additional informal information source that risk perceivers need.

Therefore, high risk perceivers will benefit more from a useful review page than low risk perceivers,

because high risk perceivers are more dependent on the reviews. If presentation format and options for

further research are contributing to review page usefulness, high risk perceivers will benefit more from

this increase in usefulness than low risk perceivers. It is therefore expected that people with a high

product risk perception of products will value the usefulness of these product review attributes more

than people with lower product risk perception. The following hypotheses are therefore proposed:

18 H4a: Product risk perception positively influences the positive relationship between sophistication of presentation format and review page usefulness.

H4b: Product risk perception positively influences the positive relationship between the option

“see more 1 2 3 4 5 star reviews” and review page usefulness.

On an individual review level, consumers tend to actually read and respond to written reviews instead of average summary statistics (Chevalier & Mayzlin, 2006). Also Archak et al. (2011) have found that, above volume and star ratings of reviews, the textual content of product reviews is an important determinant of consumer choices. People with a high product risk perception will try to decrease their insecurity by searching for additional information and could therefore be less affected by aggregated information. It could also be expected that the reliance on actual written content will be enhanced. High risk perceivers will likely be willing to sacrifice more cognitive effort to examine the product reviews and therefore rely less on aggregated review information attributes. The following hypothesis is therefore expected:

H4c: Product risk perception has a negative influence on the positive relationship between aggregated review information (Rating, Dispersion, and Summary) and review page usefulness.

2.3 Conceptual model

Based on the existing literature and proposed hypotheses, a visualization of expected relationships is shown in figure 1.

H4a H4b H4c

H2a H2b H2c

Sophistication of Presentation Format

Aggregation of Review Information

Options for Further Research

Review Page Usefulness Product Risk Perception

Figure 1: Conceptual model

19

3. RESEARCH DESIGN

In this chapter is explained which methods are used to collect the data, how the data in this study was collected and the way the data was dealt with thereafter. This chapter also provides an overview of manipulations, scales, control variables, model estimation and reliability measures.

3.1 Methodology

A conjoint analysis has been conducted in order to collect data about preferences from respondents.

More specifically, a choice-based conjoint analysis (CBC) has been used. Conjoint analysis is a multivariate technique developed specifically to understand how respondents develop preferences for any type of object; products, services, websites, etc. (Hair, Black, Babin, Anderson, & Tatham, 2010).

Since the number of attributes are relatively small <6 and choice-based conjoint enables to do analysis on both aggregate and individual level, choice-based conjoint analysis was found to be appropriate for this study. In the choice-based conjoint analysis in this study, respondents are forced to choose between three profiles. The optimum combination of attribute levels is being determined as well as the contribution of each attribute and attribute level to review page usefulness. Also, interaction effects can be determined by this methodology which is interesting for determining the effects of aggregated review information (Rating, Dispersion and Summary). Finally, groups or segments are being defined based on the preferences of individual respondents.

Segments were determined by examining the relative fit of the model, measured by the Consistent Akaike Information Criterion (CAIC). The lowest CAIC score indicates the ideal number of segments.

Besides CAIC, classification errors and segment sizes have been taken into account.

3.1.1 Model specification

Utility, the measure of respondent’s value on each attribute, is estimated with the following model:

Where Utility (U) of all the respondents of a review page (j) is the sum of all utilities (β) of the

explanatory variables (X; k=1,..,K).

20

3.2 Stimuli

Five different attributes are selected for the choice-based conjoint analysis; presentation format, options for further research, rating, dispersion and summary. The attribute presentation format contains three levels; “date” represents a presentation format whereby reviews are shown based on date (most recent review first), “usefulness” represents a presentation format whereby review are shown based on usefulness ratings of individual reviews (most useful rated first) and “usefulness + rating” represents a presentation format whereby the most useful rated positive review is contrasted with the most useful rated negative review.

The aggregation attributes Rating, Dispersion and Summary are included as separate attributes in order to prevent respondents from automatically choosing the alternative with the most information. The five attributes represent important elements of a review website on an aggregated level. Presentation format and options for further research have three levels whereas rating, dispersion and summary have two levels. Therefore, 3x3x2x2x2=72 different profiles could be created. An overview of the attributes and attribute levels are shown in table 1.

A pictorial presentation of the presentation attributes has been used in this study. A pictorial design presents the stimuli more realistic. Information overload is reduced because respondent don’t have to visualize all the information. The homogeneity of perceptions of the presentation attributes is higher among respondents and the task itself is more interesting for respondents (Green & Srinivasan, 1978).

The designs of the profiles were created in Microsoft Publisher and were based on the most commonly

Attributes Levels

Presentation format Date Usefulness

Usefulness + Rating

Options for further research See more 1 2 3 4 5 star reviews See more reviews of this reviewer See more reviews of similar products

Rating Yes

No

Dispersion Yes

No

Summary Yes

No

Table 1: Attributes and levels

21 used designs used by amazon.com, bol.com, and Yahoo.com. However, the designs have been made as simple as possible to avoid confusion and information overload. The product that was selected for the conjoint analysis was a television. The television serves as a reference to give people an idea of a real purchase situation. A television is a product that all respondent are familiar with and has been successfully used in conjoint settings before (Antilla, van den Heuvel, Rob R, & Moller, 1980; J. Lee, Cho, Lee, & Lee, 2006; Okechuku, 1994). Moreover, a television is a product that in general people tend to write a review about. On amazon.com the first page of televisions (16 televisions) comprise more than ten thousand single product reviews.

3.3 Design

The CBC contained two versions with each eight choice tasks and one hold-out task. Each task contained three profiles. The purpose of the study is to measure page usefulness. Since respondents were asked to select the most usefulness page design instead of a purchasing decision, a none-option is not included in the design. The hold-out question was identical in each of the two versions. Full profiling is used in order to keep the conjoint realistic and reduce the number of comparisons. The seventeen choice sets (eight questions v1 + eight questions v2 + one hold-out question) contained the different stimuli as described in table 1. The content of the conjoint analysis has been determined with Sawtooth Software. Due to the size and scope of this research, two fixed designs have been created resulting in nine conjoint questions with three choice tasks per respondent. Although respondents are able to answer up to 20 conjoint tasks (Johnson & Orme, 1996), two designs have been chosen above one fixed design in order to make the surveys not longer than necessary. Respondents will be less annoyed and are likely to be better able to keep their attention. The complete enumeration method is used for creating the content of the conjoint. All possible attribute levels were randomly assigned but appeared an equal amount of times in the conjoint. Each level was shown a minimum number of times in a single conjoint task and the attributes were chosen independently of other attributes. The conjoint tasks were therefore as different as possible. An overview of the attributes and efficiencies based on two versions with each eight conjoint questions are shown in table 2. Each of the levels of presentation format and options for further research are shown 16 times, the levels of rating, dispersion and summary are shown 24 times.

This results in efficiencies of 0.9567 and higher for all the levels. The design can therefore be regarded

as nearly optimal. The efficiencies are generated on the assumption that every conjoint version has an

equal amount of respondents.

22

Attribute Level Freq. Actual Ideal Efficiency

Presentation format Date 16

Usefulness 16 0.3542 0.3536 0.9961 Usefulness + Rating 16 0.3543 0.3536 0.9960 Options for further research See more 1 2 3 4 5 star reviews 16

See more reviews of this reviewer 16 0.3564 0.3536 0.9840 See more reviews of similar products 16 0.3548 0.3536 0.9930

Rating Yes 24

No 24 0.3104 0.3062 0.9730

Dispersion Yes 24

No 24 0.3130 0.3062 0.9567

Summary Yes 24

No 24 0.3079 0.3062 0.9888

Table 2: Choice-based conjoint design

However, because the sample size (respondents x conjoint tasks) and the number of questionnaire versions are relatively small, the design has been tested. Using Sawtooth Software, a virtual sample of 230 respondents randomly answered the questions. The results are shown in table 3. The standard errors of all the attribute levels are below 0.05. Also the standard errors of the interaction effects of rating, dispersion and summary are all lower than the statistical acceptable error margin of 0.05. Based on these results it can be assumed that based on these number of questionnaire versions, the number of tasks and the number of respondents (>230) the conjoint will produce statistically valid results.

Attribute Level Std. Error Interaction Std. Error

Presentation format Date 0.03435 Rating Yes x Dispersion Yes 0.04112 Usefulness 0.03678 Rating Yes x Dispersion No 0.04112 Usefulness + Rating 0.03430 Rating No x Summary Yes 0.04112 Rating No x Summary No 0.04112 Options for further

research

1,2,3,4,5 star reviews 0.03344

Reviews of this reviewer 0.03689 Dispersion Yes x Rating Yes 0.04320 Reviews of similar products 0.03400 Dispersion Yes x Rating No 0.04320

Dispersion No x Summary Yes 0.04320 Rating Yes 0.02530 Dispersion No x Summary No 0.04320

No 0.02530

Summary Yes x Rating Yes 0.04948 Dispersion Yes 0.02629 Summary Yes x Rating No 0.04948

No 0.02629 Summary No x Dispersion Yes 0.04948 Summary No x Dispersion No 0.04948

Summary Yes 0.02579

No 0.02579

Table 3: Test design (n=230)

23

Both surveys versions contain an identical hold-out question. This question serves as a check to examine the predicting accuracy of our model and is therefore not used for estimation of the utilities. The holdout question is shown in appendix 1, question 9. The nine conjoint tasks were shown in a random order to each respondent to exclude biases resulting from order, learning or annoyance.

3.3.2 Interaction effects

Interaction effects have been estimated using four additional dummy variables. Three interaction combinations contained the presence of two attribute levels (Rating + Dispersion, Rating + Summary and Dispersion + Summary) whereas the fourth interaction variable contained the presence of all three aggregated review information attributes (Rating + Dispersion + Summary). A combination was marked as a 1, when a combination was present in a profile, 0 otherwise. The interactions have been included as separate attributes to estimate the model.

3.4 Product risk perception

The moderator product risk perception has been measured with validated items of multiple scales used

in literature. The items of product risk perception were taken from different scales for a number of

reasons. First, the majority of the scales were measuring risk perception instead of product risk

perception (Choi & Lee, 2003; Forsythe & Shi, 2003; Garbarino & Strahilevitz, 2004). Among the multiple

risk perception items of these scales, the items in these scales that specifically measured product risk

perception were extracted. Second, validated scales were aimed at an offline setting making them

inappropriate for this study (Forsythe, Liu, Shannon, & Gardner, 2006). Items as “inability to touch and

feel the item” do not apply to an online setting. Third, items were product type specific: “can’t try on

clothing online” (Forsythe, Liu, Shannon, & Gardner, 2006). Therefore four items from different scales

were combined to measure the construct of product risk perception in an online setting. An overview of

the questions, sources and scales can be found in table 4. All the items are measured on a 7-point Likert

scale ranging from “absolutely disagree to absolutely agree”. An exploratory and a confirmative factor

analysis have been conducted to check the internal consistency of the product risk perception concept.

24

Question Source Scale

Product Risk Perception

1. I find it difficult to judge quality of products. (Forsythe & Shi, 2003) 1-7 Likert scale

Absolutely disagree – Absolutely agree

2. I am afraid products do not perform as expected. (Garbarino & Strahilevitz, 2004)

1-7 Likert scale

Absolutely disagree – Absolutely agree

3. I am confident that delivered products match those on the website. *

(Choi & Lee, 2003) 1-7 Likert scale

Absolutely disagree – Absolutely agree

4. I find it annoying that I can’t examine the actual products. (Forsythe et al., 2006) 1-7 Likert scale

Absolutely disagree – Absolutely agree

*Question requires reversed scoring

Table 4: Validated questions Product Risk perception and OCR Usefulness

3.5 Control variables

In order to describe the sample and characterize segments later in this research, several control variables are incorporated in this research. The following demographic variables are included: gender, age, education, profession, income and location. Besides the demographic questions, several questions are included to verify the respondents experience with internet and online shopping. Therefore, internet experience, online shopping experience and number of online sales per year are included in this research.

3.5 Procedure

All respondent received a link, which redirected them to the online survey .The respondents were asked

to imagine a situation in which they were about to buy a television. Next, before making any decisions

they were redirected to the conjoint analysis and asked to look at the different profiles carefully. The

respondents were randomly assigned to one of the two versions and also the questions were

randomized in order to filter out any learning effects. Then, the respondents were asked which of the

conjoint profiles was most useful for their purchase decision. The complete survey can be found in

appendix 1.

25

4. RESULTS

In this section the results of the conjoint analysis are being presented. First, descriptive statistics are shown. Second, the exploratory and confirmative factor analyses are performed. Third, attributes and level preferences are estimated. Fourth, interaction effects are shown. Fifth, the predictive validity of the model is evaluated. Finally, segments and segment preferences are distinguished.

4.1 Descriptive statistics

232 respondents have participated in this study of which 55.2% was male and 44.8% was female. This number is sufficient to estimate all attributes and interaction effects with a standard error lower than 0.05 (see section 3.3 design). The majority of respondents was student (65.9%) or employed (31.9) and had a HBO (24.6%) or WO education (69.4%). The average age of the respondents was 25 years old and ranged from 14 to 62 years. The majority of incomes was below € 500, - (30.6%) or between € 501, - and

€1.500, - (36.2%). The North of NL was the area in the Netherlands where most of the respondents were living (56.9%). An overview of all the descriptive statistics is shown table 5.

Gender Freq. Perc. Education Freq. Perc. Profession Freq. Perc.

Male 128 55.2 Secondary school 3 1.3 Scholar 2 0.9

Female 104 44.8 VMBO/MAVO 2 0.9 Student 153 65.9

MBO 9 3.9 Employed 74 31.9

HBO 57 24.6 Unemployed 2 0.9

WO 161 69.4 Retired 1 0.4

Total 232 100 Total 232 100 Total 232 100

Location Freq. Perc. Income Freq. Perc. Age Freq Perc.

North of NL 132 56.9 ≤€ 500 71 30.6 ≤14 1 0.4

East of NL 46 19.8 € 501 ≥ €1.500 84 36.2 15≥22 74 31.9

South of NL 10 4.3 € 1.501 ≥ € 2.500 42 18.1 23≥30 133 57.3

West of NL 25 10.8 € 2501 ≥ € 3.500 14 6.0 31≥38 8 3.4

Mid NL 12 5.2 € 3.501 ≥ € 4.500 4 1.7 39≥46 5 2.2

None of above 7 3.0 € 4.501 ≥ € 5.500 2 0.9 47≥54 5 2.2

>€ 5.501 3 1.3 ≥55 6 2.6

Rather not say 12 5.2

Total 232 100 Total 232 100 Total 232 100

Table 5: Descriptive statistics

26 Besides the demographic variables, internet experience, online shopping experience and the amount of online purchases per year are also insightful metrics. The variables are graphically shown in figure 2. The majority of the respondents make 1-12 online purchases every year. Outliers in the data purchased around 50 times online per year and interestingly, none of the respondents answered the question with 0. Thus, each of the respondents buys at least once a year a product online. Around 95% of the

respondents is considering themselves as experienced internet users (answers 5-7). Online shopping experience is less distinct. However, still the majority (71%) of respondents is considering themselves as experienced in online shopping (answers 5-7).

4.2 Scale validation product risk perception

Product risk perception has been measured using standards items from different validated scales. In theory, the four items for product risk perception measure the same concept. In order to measure if this assumption holds for the selected items of product risk perception, an exploratory and a confirmatory factor analysis have been conducted. For both analyses, the answers on the reversed scored question were turned.

First the exploratory factor analysis using a maximum likelihood method has been performed. Based on the Kaiser-Meyer-Olkin (KMO), it was appropriate to factorize the data. Product risk perception (0.701)

10 20 30 40 50 60 70 80

1≥4 5≥8 9≥12 13≥16 17≥20 ≥20

Frequency

Online Purchases per year

0 20 40 60 80 100 120

Absolutely not

3 5 Absolutely

Frequency

Experience

Internet Experience Online Shopping Experience

Figure 2: Online purchases Figure 2: Online purchases and Internet experience

27 is significantly higher (p=0.000) than the in general required level of 0.5 (Malhotra, 2010). When tested for internal consistency, the concept shows a Cronbach’s Alpha that exceeds 0.6. The Cronbach’s Alpha of product risk perception (0.706) is therefore statistically sufficient to conclude that the used product risk perception items in this study can be used for factorization (Malhotra, 2010). An overview of the KMO, significance level and Cronbach’s Alpha score are shown in table 6. A one factor solution is selected because the initial eigenvalue of the second factor is below 1 (0.844) (Malhotra, 2010). The initial eigenvalues and factor scores of all 4 items are shown in table 6. The factor loading scores have a reasonable magnitude.

Exploratory factor analysis Confirmatory factor analysis

KMO Sig Cronbach’s Alpha Chi-sq Df Chi-sq/Df p- value RMSEA

Product risk perception 0.701 0.000* 0.706 5.01 2 2.5005 0.08164 0.081

Initial eigenvalues GFI AGFI NFI CFI

Factor Tot. % of Var. Cum. % 0.99 0.95 0.98 0.99

1 2.139 53.487 53.487

2 0.844 21.105 74.592

3 0.59 14.75 89.342

Item Factor loading scores Factor loading scores

Prod. Risk Perception Q1 0.638 0.64

Prod. Risk Perception Q2 0.832 0.83

Prod. Risk Perception Q3 0.422 0.42

Prod. Risk Perception Q4 0.579 0.58

*Significant on a 99% significance level.

Table 6: Factor analyses product risk perception

Second, a confirmatory factor analysis has been conducted to test the construct validity. Construct

validity refers to the degree to which product risk perception related to the items (Hoyle, 2000). The

results are shown in table 6. A chi-square with 2 degrees of freedom (Chi-sq/Df = 2.5005) is fairly good

(ideally below 2.0). The p-value is 0.082, exceeding the required 0.05 (Hoyle, 2000). Also the root mean

square of residuals (RMSEA) is fairly good (0.081) where ideally below 0.05 (Hoyle, 2000). The goodness

of fit indicators GFI (0.99), AGFI (0.95), NFI (0.98) and CFI (0.99) all exceed the required level of 0.9 and

indicate a good fit (Hoyle, 2000). The third item has a relatively lower factor loading score (0.422) in

both analyses. This item is however included in the model because some of the disturbance may be

caused by the reversed coding of the item and not necessarily by the item itself. Based on the

exploratory and confirmatory factor analysis it is concluded that the four items measure product risk

perception on a satisfactory level. A sum variable of the four items has been created to represent

product risk perception in the conjoint analysis.

28

4.3 Conjoint analysis

Since all of the attributes are nominal, part-worths are being estimated in order to determine the relative importance of the attributes on review page usefulness for all the respondents. An overview of the estimated part-worths, p-values, ranges and relative importance of the attributes is shown in table 7. All the part-worths of the attribute levels are significantly different from each other, represented by a p-value of 0.000 for all of the attributes. Summary is the most important contributor to review page usefulness (31%), followed by rating (26%), Dispersion (19%), Presentation format (13%). Options for further research was the least important contributor to review page usefulness (11%).

Attribute Level Estimated

Part-worth

Wald p-value Range Relative Importance

Presentation format Date -0.2040* 50.79 0.000* 0.4565 13%

Usefulness -0.0485

Usefulness + Rating 0.2525*

Options for further research

See more 1 2 3 4 5 star reviews 0.0874* 34.60 0.000* 0.3877 11%

See more reviews of this reviewer -0.2376*

See more reviews of similar products 0.1502*

Rating Yes 0.4712* 241.22 0.000* 0.9423 26%

No -0.4712*

Dispersion Yes 0.3503* 149.61 0.000* 0.7006 19%

No -0.3503*

Summary Yes 0.5583* 315.13 0.000* 1.1166 31%

No -0.5583*

*Significant on a 95% significance level, z-value > 1.96 Table 7: Attribute and level parameters

Based on the information in table above, several graphs can be created to show the influence of the

different attributes on review page usefulness. First, in figure 3, the relative importance of the different

attributes is shown. Second, the direction of the attribute levels is presented.

29 A more sophisticated presentation format has a higher utility compared to a less sophisticated presentation format. A presentation format based on date has the lowest (negative) utility (-0.204) followed by a presentation format based on usefulness (-0.0485). A presentation format whereby the most useful rated positive review is contrasted with the most useful rated negative review has the highest utility (0.2525). The option: “See more reviews of similar products” has the highest utility (0.1502), followed by “See more 1 2 3 4 5 star reviews” (0.0874). “See more reviews of this reviewer”

has the lowest utility (-0.2376). Showing Rating (0.4712), Dispersion (0.3503) and Summary (0.5583) all strongly contribute to review page usefulness whereas not showing them; Rating (-0.4712), Dispersion (- 0.3503) and Summary (-0.5583) has a strong negative effect on review page usefulness.

13%

11%

26%

19%

31%

Relative importance of attributes

Presentation format Options for further research Rating

Dispersion

Summary

-0,3 -0,2 -0,1 0 0,1 0,2 0,3

Date Usefulness Usefulness + Rating

Parameter

Presentation format

-0,3 -0,25 -0,2 -0,15 -0,1 -0,05 0 0,05 0,1 0,15 0,2

See more 1 2 3 4 5 star reviews

See more reviews of this reviewer

See more reviews of similar products

Parameter

Options for further research

-0,8 -0,6 -0,4 -0,2 0 0,2 0,4 0,6 0,8

Yes No

Rating

Parameter

Rating, Dispersion and Summary

Rating Dispersion Summary

Figure 3: Relative importance of attributes and directions of attribute levels

30

In the previous sections, the relationship between aggregated review information (Rating, Dispersion and Summary) and review page usefulness has been confirmed. It is interesting to examine if interaction effects between the aggregated review attributes exist. In other words: Is the utility of showing for example rating and dispersion together bigger than the sum of the individual utilities of rating and dispersion? Considerable evidence has been found that people treat attributes as complementary (Wittink & Cattin, 1989). This study is examining the importance of the aggregated review information attributes by showing vs. not showing them. Complementarities among attributes could therefore very well exist. Moreover, interaction effects are more likely to occur in pictorial than verbal representations (Wittink, Vriens, & Burhenne, 1994). Interaction effects could be estimated by incorporating new interaction attributes to the existing model (Green & Srinivasan, 1978).

Four dummy product attribute variables were created in order to capture the interaction effects.

Attribute variable one, two and three (Rating + Dispersion, Rating + Summary and Dispersion + Summary) were marked as a 1 when the conjoint profile contained that specific combination of two of the aggregated information attributes and 0 if this was not the case. The fourth variable was marked 1 when all three aggregated review information attributes were present in the profile, 0 otherwise. The combinations were respectively 5, 6, 6 and 4 times present in the conjoint. The results of the parameter estimates when the interaction effects are included in the existing model are shown in table 8.

Dispersion in combination with a summary has a significant positive interaction effect on review page usefulness (Wald=3.85, p=0.05). All other interactions (Rating + Dispersion, Rating + Summary and Rating + Dispersion + Summary) are insignificant (p-value>0.10).

Attribute Present in profiles Parameter Z-value Wald P-value

Rating + Dispersion

5x 0.1084 0.4353 0.1895 0.660

Rating + Summary

6x -0.0231 -0.0821 0.0067 0.930

Dispersion + Summary

6x 0.5948 1.9623 3.8506 0.050*

Rating + Dispersion + Summary 4x

-0.0297 -0.0671 0.0045 0.950 Significant on a 0.90% significance level

Table 8: Parameter estimates interaction effects

31

4.4 Hypotheses testing

A more sophisticated presentation format has a significantly higher contribution to review page usefulness as shown in table 7 and figure 3. Showing the most useful positive review contrasted with the most useful negative review represents the highest usefulness. A presentation format showing the most useful reviews has higher utility than a review format based on date. Based on these results, H1 is supported. H2a, H2b and H2c are also supported. Showing aggregated review information versus not showing the information significantly contributes to review page usefulness. Showing a rating (vs. not showing rating) has a significantly positive effect on usefulness. The inclusion of dispersion and/or summary both significantly increase the usefulness of online reviews (p=0.000). The graph is showing a clear downward slope for each of the aggregated information criteria. Moreover, interaction effects exist between dispersion and summary. The utility of showing a combination of these information attributes is higher than the sum of the individual utilities.

The option “See more 1 2 3 4 5 star reviews” significantly differs from the other options for further research. Despite the fact that the option “See more 1 2 3 4 5 star reviews” has a positive utility, the option has not the highest utility of all three options. The option “See more reviews of similar products”

is contributing more to review page usefulness. Therefore, hypothesis H3 is rejected. In sum, showing online reviews using a sophisticated presentation format whereby the most useful rated positive review is contrasted with the most useful rated negative review, displaying all three aggregated review information criteria (Rating, Dispersion and Summary) and the option “See more reviews of similar products” represents the highest review page usefulness.

4.5 Predictive validity

The predictive validity of the model can be estimated internally and externally. Latent Gold provides an internal prediction tool using 1856 replications. The results are shown in table 9. The hit rate of the internal prediction tool is 73.1%. In 73% of the cases, the model predicted right.

All 232 respondents have answered nine conjoint questions. One of these questions (hold-out question)

has not been used for estimation of the model. Instead, this question is used to verify the external

predictive validity of the model. The predicted choice for each respondent was the profile with the