HOW REVIEWS ARE REVIEWED: THE IMPACT OF
READABILITY AND SOCIAL FACTORS ON THE USEFULNESS
OF ONLINE CONSUMER REVIEWS
HOW REVIEWS ARE REVIEWED: THE IMPACT OF
READABILITY AND SOCIAL FACTORS ON THE USEFULNESS
OF ONLINE CONSUMER REVIEWS
Mariska Verstappen University of Groningen Faculty of Economics and Business
MSc Marketing Intelligence January 16, 2014 Bataviastraat 23a 9715KJ Groningen Tel: +31(0) 651660457 E-mail: m.verstappen@student.rug.nl Student number: S1879219 Supervisors University of Groningen
Management summary
Much prior research in the field of online consumer reviews has focused on the positive effect of reviews on both sales and conversion. Now that the use of online consumer reviews has become more widespread, the focus shifts from the mere presence of reviews to how these reviews are actually processed by consumers. Given the importance of reviews on conversion and the large amount of reviews that is often available, is it important that online retailers carefully consider how the reviews should be presented on their website.
Therefore, this research makes a start by studying how reviews are reviewed. This differs between consumers and therefore three segments with different preferences are distinguished. These are labelled ‘structure seekers’, ‘confirmation seekers’ and ‘self-determined customers’. Moreover, in this study, six different attributes that were expected to influence review usefulness are compared.
Abstract
Online consumer reviews have received much research attention in the last couple of years, because it has become one of the most frequently accessed online information sources. Based on previous literature, this empirical research distinguishes between six attributes that are expected to influence review usefulness. The attributes that are included in this study fall within two categories that have received only limited research attention, namely readability factors and social factors. The readability factors that are included are ‘degree of structure’ and ‘spelling errors’. The social factors that are investigated are ‘real name exposure’, ‘homophily between reader and reviewer’, ‘expert status’ and ‘usefulness others attach to the review’. Next to these attributes, ‘product category’ and ‘reader’s product expertise’ are included as moderators, to evaluate whether they influence the expected relationships. Data was collected by means of two surveys that were distributed among customers from a Dutch online retailer, including a choice-based conjoint task. To measure whether ‘product category’ has a moderating role, approximately half of the respondents evaluated reviews from the high-involvement product ‘tablet’, whereas the other respondents evaluated reviews from the medium-involvement product ‘Dutch novel’. Using this data, a choice-based conjoint analysis (on both aggregate and segment level) was performed in order to find out how important the distinct attributes are to consumers. The results show that there are no big differences regarding importance of the attributes between the product categories or between readers with different levels of product expertise. The aggregate results show that readability factors, with degree of structure in particular, are more important than social factors when it comes to review usefulness. However, except for real name exposure, all social factors affect review usefulness as well. Three different segments with different preferences are distinguished for both product categories, which are labelled ‘structure seekers’, ‘confirmation seekers’ and ‘self-determined customers’. This research brings general implications for online retailers, as well as specific implications directed to individual segments. Furthermore, this research contributes to the current literature on online consumer reviews as new variables are explored and bases for future research are provided.
Table of contents
1. INTRODUCTION ... 1
2. THEORETICAL FRAMEWORK ... 5
2.1. READABILITY FACTORS ... 5
Spelling errors ... 5
Degree of structure ... 6
2.2. SOCIAL FACTORS ... 6
Real name exposure ... 7
Homophily between reader and reviewer ... 8
Expert status of reviewer ... 8
Usefulness others attach to the review ... 10
2.3. MODERATORS ... 11
Information processing and the elaboration likelihood model ... 11
Product category (high involvement versus medium involvement products) ... 11
Reader’s product expertise ... 13
2.4. CONCEPTUAL FRAMEWORK ... 14
3. METHODOLOGY ... 15
3.1. METHOD ... 15
Choice-‐based conjoint analysis ... 15
3.2. DATA COLLECTION ... 16
Sampling ... 16
Measurement scales ... 16
Attributes and levels ... 18
3.3. STUDY DESIGN ... 20
3.4. PLAN OF ANALYSIS ... 22
Construct validation ... 22
Independent Samples T-‐tests ... 23
Model specification and estimation ... 23
Construct validity ... 28
Post-‐tests on importance of reviews and purchase decision involvement ... 28
4.1. CBC ANALYSIS FOR THE MEDIUM-‐INVOLVEMENT PRODUCT ‘DUTCH NOVEL’ ... 29
Aggregated model ... 29
Model selection ... 31
Segment interpretation ... 32
Validity ... 35
4.2. CBC ANALYSIS FOR THE HIGH-‐INVOLVEMENT PRODUCT ‘TABLET’ ... 35
Aggregated conjoint analysis ... 35
Model selection ... 37
Segment interpretation ... 38
Validity ... 40
4.3. COMBINED AGGREGATE MODEL TABLETS AND DUTCH NOVELS ... 41
4.4. HYPOTHESIS TESTING ... 42 5. DISCUSSION ... 43 5.1. THEORETICAL IMPLICATIONS ... 44 Readability factors ... 44 Social factors ... 45 Moderators ... 46 5.2. MANAGERIAL IMPLICATIONS ... 47
5.3. LIMITATIONS AND DIRECTIONS FOR FUTURE RESEARCH ... 49
1. Introduction
As consumers become more aware of the influence of marketers in traditional information channels (Godes & Mayzlin, 2004), online consumer reviews have become one of the most frequently accessed online information sources (Ludwig, De Ruyter, Friedman, Bruggen, Wetzels & Pfann, 2013). Online consumers reviews are perceived to be more credible and trustworthy than producer-generated content (Purnawirawan, De Pelsmacker & Dens, 2012).
A lot of research is done that has shown the effects of online consumer reviews on company performance. These studies show that the presence of online consumer reviews has both a positive effect on conversion (Ludwig et al., 2013) and a positive effect on sales (Chevalier & Mayzlin, 2006; Zhu & Zhang, 2010). However, as the availability of online consumer reviews on websites has become widespread, the focus now shifts from the mere presence of reviews, to how these reviews are processed and evaluated (Mudambi & Schuff, 2010). In the last couple of years, researchers have become more interested into what aspects of a review makes that a review is more or less useful in the eyes of consumers (Cao, Duan & Gan, 2011; Purnawirawan et al., 2012).
While a consumer reads a message, he or she assigns a certain level of usefulness to the message that is critical for the adoption of the information that is presented. The information that is not perceived to be useful will not be taken into consideration (Purnawirawan et al., 2012). The volumes of the available online consumer reviews, as well as the great variations in quality are big obstacles to consumers who want to make use of reviews prior to their purchase (Liu et al., 2008). For example, the amount of reviews on Amazon for an average book can easily be more than hundred. For some very popular books, the amount of reviews can even be in the thousands. In those cases, it becomes very unlikely that a consumer reads all the reviews prior to his purchase. What a consumer really needs could be just a subset of the most useful reviews (Cao, Duan & Gan, 2010). It is important to get deeper insights into the factors that make one review more useful than the other. In that way, general ways of improving the review system and other bases for sorting the reviews in an online shop might be detected.
the perceived usefulness of a review (Willemsen, 2011). A lot of these prior studies focus on content variables. Although these studies are all useful contributions to the literature on online consumer reviews, most of these factors are hard to deal with in practice, as online retailers cannot change the content of reviews that people post.
In this research, two groups of factors are studied that are more manageable for online retailers, namely social and readability factors. Both of these topics are underexplored in the literature on online consumer reviews.
Although recent research focussed on content and linguistic style matches (Ludwig et al., 2013), research focusing on readability factors in reviews is still very scarce. Readability did receive much attention in other areas, like in billboard advertising and direct mailing. In those contexts, readability has shown to influence the processing attention that a message receives (Hozier & Robles, 1985; Taylor, Franke & Bang, 2006; Sherman, Greene & Plank, 2011). Since online consumer reviews have gained so much influence recently (Ludwig et al., 2013), it is interesting to get more knowledge about the impact of readability factors in the context of online consumer reviews.
Social cues in online consumer reviews also received only limited attention. While previous studies gave a lot of interesting insights, they focus mostly on informational, and not on social processes that might play a role in the adoption of review content. However, as online consumer reviews are essentially consumer-to-consumer communications, it is logical to assume that both informational as well as social factors do play a role in the adoption of information (Racherla & Friske, 2012). Although some social concepts have received
attention in the last couple of years(Forman, Ghose & Wiesenfield, 2008; Baek, Ahn & Choi,
2013), it would be interesting to know how much influence social factors actually have on review usefulness, when they are directly compared to other factors. In this study, social factors are compared with readability factors.
For readability, the degree of structure and the amount of spelling errors will be explored. In this study, the structure of the review is explained in terms of simplicity with which the text is presented, spacing and the use of clear headlines in texts. To the author’s knowledge, the effect of review structure on review usefulness has never been empirically explored before.
before, the impact of usefulness others attach to the review on review usefulness has never been explored empirically, to the author’s knowledge.
When consumers evaluate online consumer reviews, there are different processing strategies they can use. The elaboration likelihood model (ELM) explains these strategies in terms of central cues and peripheral cues in information processing (Petty & Cacioppo, 1983). In this article, it is reasoned that social factors reflect peripheral cues and readability factors affect the central route in information processing. When people have a high ability and motivation to process the information, they are more likely to follow the central route of information processing, whereas low scores on these factors make them more likely to follow the peripheral route (MacInnis, Moorman & Jaworski, 1991). In this study, ability is reflected by a reader’s product knowledge, whereas motivation is reflected by product category (high-involvement vs. lower (high-involvement products). These are included as potential moderators in this study. It is expected that readability factors are more important for high involvement products and for readers with more product knowledge. At the same time, it is expected that social factors are more important for lower involvement products and for readers with less product knowledge. In short, the aim of this paper is to answer the following research question:
“To what extent do readability and social factors affect the usefulness of reviews among different product categories and among people with different levels of product expertise?”
In this study, data is gathered by means of a survey among customers from a Dutch online retailer. The data will be analyzed by means of a Choice-Based Conjoint analysis, to investigate the extent to which the different readability and social attributes affect review usefulness on both aggregated and segment level.
Results show that readability factors are more important than social factors when it comes to review usefulness, with degree of structure being particularly important. Besides, all social factors, except for real name exposure, affect review usefulness as well. No big differences regarding importance of the attributes between the product categories or people with different levels of product expertise were found. Three different segments with different preferences are distinguished for both product categories, which are labelled ‘structure seekers’, ‘confirmation seekers’ and ‘self-determined customers’.
are underexplored. Furthermore, not much research has been done that directly compares the extent to which different factors affect review usefulness. The Choice-Based Conjoint technique that is used in this study makes it possible to investigate this since consumers have to make direct trade-offs between different attributes. These differences in importance between attributes are investigated not only on aggregated level, but also on segment level. Furthermore, two variables that are included in this study have, to the author’s knowledge, never been explored empirically in the context of online consumer reviews. These variables are degree of structure and usefulness others attach to the review. Hence, the literature on online consumer reviews gets a valuable extension, as new interesting concepts will be explored.
What makes this study even more interesting is the strong practical relevance of the attributes that are explored. A lot of prior studies concerning online consumer reviews have primarily focused on content variables. Although these studies give interesting theoretical insights, they have less practical relevance since online retailers cannot change the content of the reviews that customers post on their website. This study, in contrast, explores attributes that are manageable for online retailers and gives online retailers clear insights on how to make their review pages as useful as possible for their customers, on both aggregate and segment level.
The remainder of this paper is structured as follows. First, there will be a literature review where the readability and social factors are elaborated and hypotheses are formed. Together with the expected moderators, these will be summarized in a conceptual framework. The next parts are the methodology and results section, where the Choice-Based Conjoint analysis is discussed and the preferences of consumers regarding the attributes are analyzed. In the final section, the results are discussed and implications for online retailers are given, as well as limitations and directions for future research.
2. Theoretical framework
The aim of this article is to find out to what extent readability and social factors affect the usefulness of reviews. The various readability and social factors that are included in this study are elaborated on in this section. First, spelling errors and degree of structure are discussed and readability hypotheses are formed. Second, the social factors (real name exposure, homophily between reader and reviewer, expert status and the usefulness other attach to the review) are elaborated on. Third, the elaboration likelihood model is discussed and the expected moderators product category and a reader’s product expertise are elaborated on. This section is summarized by means of a conceptual model.
2.1. Readability factors
Readability refers to the linguistic complexity of a text. It centres on words, sentences and text density (Chebat, Charlebois, & Gélinas-Chebat, 2001). In other words, readability refers to how easy it is to read a body of text (Conti-Ramsden, Durkin & Walker, 2012).
If a message is difficult to understand, the reader is likely to infer negative thoughts about the communicator (Schindler & Bickart, 2012). Thus, it is expected that the presence of style variables that reduce the understandability of messages will be associated with less useful reviews. In this research, two readability factors will be explored.
Spelling errors
Spelling errors affect the readability of a message. Spelling errors within a message can lead to misunderstanding, ambiguity and difficulties in the comprehension of a message (Sallis & Kassabova, 2000; Ghose & Ipeirotis, 2011). Jessmer & Anderson (2010), show that people who wrote grammatically correct email messages were perceived to be more competent and more likable than people who wrote grammatically wrong messages. Prior research already investigated the importance of the correctness of spelling in the context of reviews and found that spelling errors indeed reduced the perceived usefulness of reviews (Forman, Ghose & Wiesenfield, 2008; Ghose & Ipeirotis, 2011; Schindler & Bickart, 2012). However, the relative importance of spelling errors compared to other review factors is never explored empirically. This research applies a methodology that makes it possible to explore the relative impact of spelling errors on review usefulness, to see how important this factor is compared to other factors that are included in the study. The following hypothesis will be tested:
Degree of structure
Beard & Williams (1988) studied readability measures in direct mail campaigns. They suggest that the letter must be appropriate for the target. For example, it must reflect the target’s linguistic repertoire, like familiarity with technical terms. A limitation of this readability study is that it does not tell how to format or organize the text.
Sherman et al. (1991) investigated the effectiveness of three different message structures in the context of direct mail campaigns. They found that message structure matters and argue that companies must test multiple message structures in identifying whether specific formats can lead to higher response rates. Hozier & Robles (1985) also found that the structure of the letter appears to have some effect on the level of interest in direct mailing campaigns.
In the context of online consumer reviews, companies do not have a direct influence on the specific arguments in the reviews. They are able, however, to set the structure in which the information is presented. As different structures have a different impact in response rates of direct mailing campaigns (Sherman et al., 1991), it is likely that structure also matters in online consumer reviews.
Taylor, Franke & Bang (2006) use several measures for advertisement text readability, which include the simplicity with which the message is presented, presenting well-spaced text and using headlines that stand out in the advertisement. So, adding structure to texts, in the form of spacing and clear headlines to texts can increase readability. Information is more likely to get processing attention when it is easily readable (MacDonald-Ross, 1977; Tufte, 1983). Therefore, the following hypothesis will be tested:
H2: More structure has a positive effect on review usefulness.
2.2. Social factors
consumer review, is uncertainty about product characteristics, whereas the second type of uncertainty has to do with the intentions the communicator has and what kind of person the communicator is (Racherla & Friske, 2012). This second type of uncertainty has to do with social factors of the message. As online consumer reviews are essentially C2C-interactions, it is thus logical to assume that both informational as well as social factors do play a role in the adoption of information (Forman et al., 2008; Racherla & Friske, 2012).
Prior research gives reason to believe that the amount of social information provided is important in consumer’s buying decisions (Baek et al., 2013; Racherla & Friske, 2012). However, there are some specific social factors that deserve more attention. In this research, the effects of four social factors on review usefulness are explored.
Real name exposure
Source credibility has shown to play a significant role in the adoption of online information (Briggs, Burford, De Angeli & Lynch, 2002; Brown, Broderick & Lee, 2007). Furthermore, Chow, Lim & Lwim (1995) show that source credibility has a positive effect on message credibility.
Credibility of reviews becomes a more commonly asked question, since some companies and third-parties (like contracted marketing firms) try to manipulate the reputation of their products (Jensen, Averbeck, Zhang & Wright, 2013). Besides, everybody can place reviews: from novice product users to product experts. To make purchase decisions, online shoppers must now rely on the characteristics of the reviewers in determining credibility (Jensen et al., 2013). Consumers sometimes read only reviews written by reviewers who, the consumers feel, are genuine (Racherla & Friske, 2012). One reviewer characteristic that contributes to credibility is the extent to which the real name is exposed.
Prior research suggests that readers of a message use source information, like a reviewers name, as simple shortcuts to help them reach judgments (Chaiken, 1980). When they do this, it is likely that a review from a reviewer with a real name gets a more positive judgment, as the source is considered to be more credible. Forman et al. (2008) found that showing a real name instead of a nickname has a positive effect on perceived review usefulness. In contrast, Baek et al. (2013) measured real name exposure by means of a real-name badge and found no significant effect on perceived review helpfulness.
their real name is also investigated, to see if there is a difference between reviewers that show a part of their real name, or their entire name. Thus, the following hypothesis will be tested:
H3: Real name exposure has a positive effect on review usefulness.
Homophily between reader and reviewer
Homophily refers to how similar individuals are in terms of certain attributes (Brown and Reingen, 1987). Identity disclosure of information sources helps individuals to find information from people who have much in common with themselves (Jensen et al., 2013).
Uncertainty reduction theory reasons that an individual may reduce uncertainty by choosing to communicate with other people who share similar values and social identity (Turner, Brown & Tajfel, 1979). Furthermore, the source-attractiveness model suggests that receivers are better able to identify with sources that are similar to themselves (Kelman, 1961).
Prior research in advertising supports the hypothesis that similar communicators are perceived to have more influence than dissimilar communicators (Feick and Higie, 1992). In the context of online consumer reviews, however, not much research is done yet that explores homophily between reader and reviewer. A recent study from Naylor, Lamberton & Norton (2011) reveals that reviews from either similar reviewers or from reviewers that do not show personal information are both more persuasive than reviews from dissimilar reviewers. Their study, however, does not explore the relative impact of homophily compared to other attributes. Besides, the study from Naylor et al. (2011) focused on reviews from a restaurant experience and was conducted among university students only. In order to find out whether their findings also hold for a broader range of people and among other product categories, homophily between reader and reviewer is studied here as well. Thus, the following hypothesis will be tested:
H4: Homophily between reader and reviewer has a positive effect on review usefulness.
Expert status of reviewer
has more influence on the purchase intentions of consumers than information that is provided by non-expert sources (Harmon & Coney, 1982; Lascu, Bearden & Rose, 1995).
If someone is an expert, is determined by the knowledge and competence a source has regarding the topic of interest (Gotlieb & Sarel, 1991). In online settings, however, it is harder to make an evaluation regarding knowledge and competence of the author of the information when the author’s background cannot be verified (Cheung, Lee & Rabjohn, 2008). As Brown et al. (2007) suggest, the online evaluation of expertise must be based on a reviewer’s self- claim. Research done by Eastin (2001) has shown that a message from someone that gave him-/herself an expert status was perceived to be more credible than a message from people who did not. Next to self-evaluation, an expert-status can also be assigned to a reviewer by a specific web-shop, perhaps based on the amount of orders placed in a specific product category. Several authors explained the effect of expert sources by the authority heuristic (Hu & Sundar, 2010; Tan, Swee, Lim, Detenber & Alsagoff, 2008), which is a cognitive decision rule implying that expert statements are true. This heuristic influences the process of information evaluation. Whenever an expert status is assigned to a source, the message will be positively evaluated, irrespective of the content of the message (Chaiken & Eagly, 1989).
Next to being a product expert, a reviewer can also be a general expert in the writing of reviews. Several web-shops assign status to reviewers who have written a lot of reviews with the badge ‘top-reviewer’. Sometimes, these badges also include a rank (e.g. top-100 reviewer). Prior research has shown that a higher activity level of a reviewer has a positive effect on the credibility and usefulness of reviews (Cheung, Luo, Sia & Chen, 2009; Baek et al., 2013).
H5a: If a reviewer is claimed to be a top-reviewer, the review is perceived to be more useful than when the reviewer has no status.
H5b. If a reviewer is claimed to be a product-expert, the review is perceived to be more useful than (1) when the reviewer has no status and (2) when the reviewer is claimed to be a top-reviewer.
Usefulness others attach to the review
The value other people attach to a review (by means of indicating how useful they think the review is) can influence how convincing a review is in the perception of the reader. Although some research is done about expert status and the exposure of demographic information, less research has been published that assesses the effect of normative influence from others on review usefulness. It would be interesting to explore how much the impact of usefulness ratings from others is, compared to other aspects of the reviews (Purnawirawan, 2012). Two social processes might be at stake here.
First, the process of observational learning might be at stake when people read a review and see that others find the review useful. Observational learning is a type of social interaction that is action- or behaviour based (Chen, Wang & Xie, 2010). Becker (2011) gives an example of observational learning in the context of restaurants, where a person might be heavily influenced by how many others there already are in the restaurant, even without knowing their identities and reasons for that decision. Observational learning in the context of online consumer reviews would imply that someone finds a review very useful, based on simply observing that many others also found the review useful. Even though the amount of votes does not reveal reasons why these reviews are considered to be more useful, actions of others often speak louder than words and observational learning can have a large impact on consumer decisions and evaluations (Chen, Wang & Xie, 2010).
Thus, based on the possible effects of conformity pressures and observational learning it is expected that usefulness-votes of others have a positive effect on review usefulness. The following hypothesis will be tested:
H6: The degree of usefulness others attach to the review has a positive effect on review usefulness.
2.3. Moderators
Information processing and the elaboration likelihood model
The elaboration likelihood model (ELM) gives a useful theoretical perspective on information processing by consumers regarding online consumer reviews (Petty & Cacioppo, 1983; Park, Lee & Han, 2007). ELM posits that an individual who has the motivation to process the message and who has the ability to process the message, is more likely to process information via the central route (MacInnis, Moorman & Jaworski, 1991). Individuals that lack the motivation or ability to process the message, on the other hand, are more likely to process the information via the peripheral route (e.g. using mental shortcuts). Involvement is associated with the motivation to process reviews and prior knowledge (expertise with the product) is associated with the ability to process information (Celsi & Olsen, 1988; Petty & Cacioppo, 1983). Below, the expected moderating roles of both product category (high-involvement products versus medium-involvement products) and a reader’s product expertise will be discussed.
Product category (high involvement versus medium involvement products)
high-involvement products. Therefore, in this research the decision is made to go in between by choosing product groups that are indeed different in their level of involvement, but not extremely low, as online consumer reviews are unlikely to be used in purchasing decisions for those products. Thus, a distinction is made between high-involvement products and medium-involvement products. Even though people are unique and they can differ in their level of involvement with products, the distinction is based on the general level of involvement with product categories because this is very common in the literature (Martin, 1998; Traylor & Joseph, 1984; Laurent & Kapferer, 1985) and more convenient to work with in practice.
Consumers spend more time and energy in the decision-making process of purchasing high-involvement products, whereas they spend less time and effort for lower involvement products (Richins and Bloch, 1986). The higher the involvement with a product, the greater the search for information about the product is, the more sources of information that are used and the more extensive the information processing by consumers is (Barber & Venkatraman, 1986; Petty & Cacioppo, 1983) because consumers want to maximize their satisfaction with the product (Barber & Venkatraman, 1986). In lower involvement products, people are less encouraged to seek for information and less motivated to give the message a lot of attention (Sussman & Siegal, 2003). Furthermore, when involvement with the product is lower, people are less motivated to put much cognitive effort into the evaluations of the content, but rely more on peripheral cues in evaluations, (Chaiken, 1980; Petty & Cacioppo, 1981; Petty & Cacioppo, 1983), such as the reviewer’s expert status, the reviewers name and usefulness-votes.
H7: For medium-involvement products, social factors are more influential in deciding on a review’s usefulness than readability factors.
H8: For high-involvement products, readability factors are more influential in deciding on a review’s usefulness than social factors.
Reader’s product expertise
A reader’s product expertise, the amount of prior knowledge the reader has about the issue (Sussaman & Siegal, 2003), also influences the way in which the reader processes the information. This prior knowledge has a major impact on the attention that is devoted to the message and the capacity to process and evaluate the information displayed (Alba & Hutchinson, 1987; MacInnis & Jaworski, 1989). It is also empirically demonstrated that product expertise affects the depth of the information processing (Chebat et al. 2001).
People with more expertise have more knowledge and ability to evaluate a message (Cheung et al., 2009). Therefore, consistent with the manner in which high-involvement products are processed, people with more expertise are more likely to focus on the content of a message (Suggman & Siegal, 2003). Cheung et al. (2009) show that consumers who are less knowledgeable on the review topic, are more likely to depend on peripheral cues such as source credibility while judging online product reviews. Therefore, in accordance with the reasoning of ELM, all relations that are expected for medium-involvement products are also expected for readers with lower product expertise. The same goes for high-involvement products and high product expertise. Therefore, the following hypotheses will be tested:
H9: For readers with lower product expertise, social factors are more influential in deciding on a review’s usefulness than readability factors.
2.4. Conceptual framework
Based on the theoretical framework, a conceptual model with the factors that are hypothesized to influence review usefulness is created, which is shown in figure 1.
3. Methodology
In this section, the research methodology is discussed. First, the decision to use the Choice-Based Conjoint (CBC) analysis will be explained. After that, the data collection procedure and the measurement scales are described, including the levels of the readability and social attributes. Finally, the study design and the plan of analysis will be elaborated on.
3.1. Method
Choice-based conjoint analysis
Conjoint Analysis is a technique for measuring trade-offs among multi-attributed products and services (Green & Rao, 1971; Johnson, 1974; Green & Srinivasan, 1990). More specifically, Conjoint Analysis is a “decompositional method that estimates the structure of a consumer’s preferences given his/her overall evaluations of a set of alternatives that are pre-specified in terms of levels of different attributes” (Green & Srinivasan, 1978, p. 4). Conjoint Analysis is the ideal method here because the aim of this research is to find out what attributes are most influential in determining review usefulness and what levels that attributes ideally have. In order to determine consumer preferences based on utilities, objects are considered as scores on a set of attributes, which means that the utility of an option is considered to be the product of the utilities of the attribute levels of that option (Johnson, 1974).
Only estimating an aggregate model would be unrealistic as people can differ in their preferences. Therefore, this research investigates the extent to which the different readability and social attributes affect review usefulness on both aggregated and segment level. Market segmentation presupposes heterogeneity in consumer’s preferences and choices (Green & Krieger, 1991).
traditional Conjoint Analysis (Louviere & Woodworth, 1983). Furthermore, the surveys used for a Choice-Based Conjoint Analysis are often much shorter than the surveys used for the other conjoint methodologies, which results in a smaller burden on respondents (Johnson & Orme, 1996). This reduces the chance that respondents get bored and might fill in random answers as a consequence.
3.2. Data collection
Sampling
To investigate whether the effect of the readability and social factors on review usefulness differs between product categories, two different types of surveys were needed. One survey contained reviews from a medium-involvement product, whereas the other survey showed reviews from a high-involvement product. Besides, the choice was made to lessen the burden on respondents of evaluating too many choice sets, by means of splitting the conjoint tasks of both surveys in two parts. Thus, in total, four surveys were distributed.
The surveys were created in SurveyWorld and are distributed among customers from a large Dutch online retailer. This online retailer sells a wide assortment of consumer goods. According to Hair et al (2010), a sample size of 200 respondents is sufficient. Each version of the survey is sent to 10.000 customers, assuring that a response rate of 2% per version would already be sufficient for a proper analysis. This margin is taken because the loading time of the survey was pretty long due to several routings in the survey. While the online retailer has both Dutch and Belgian customers, only customers from one of these nationalities had to be chosen for the data collection to ensure that the attribute homophily is measured correctly. The choice is made to only send the survey to Dutch customers, because the online retailer has much more Dutch customers than Belgian customers. Apart from the restriction on country, a random sample from all active customers that placed an order in the last half year was approached to participate.
Measurement scales
computer hardware and software as high-involvement products. Books, on the other hand, were not as often mentioned as high-involvement products and sometimes even as low-involvement products. Therefore, it is considered to be a medium-low-involvement product here.
To make sure that the products investigated are a good reflection of medium- and high involvement products, a pre-test on Purchase Decision Involvement (PDI) is held based on
Ratchford’s FCB scale (1987).Twenty persons from different genders and ages were asked to
answer three questions one a seven-point likert scale (1 = ‘Making one’s selection of this product is a very unimportant decision to me/a very important decision to me’, 2 = ‘The decision about which product to buy requires a little thought/a lot of thought’ and 3 = In deciding which product to buy, there is a little to lose/there is a lot to lose’). A paired-Samples T Test was used to compare the means of the PDI scores for both product groups. The results of this test show that the involvement when buying a tablet (M = 6,53, SD = 0,31), is indeed significantly higher than the involvement when buying a Dutch novel (M = 4,42, SD = 0,67, t(19): -12,46, p<0,001). To see whether the difference in involvement also holds for the respondents of the CBC studies, the PDI question is also asked to all respondents that participate in the main studies.
To measure the second moderator, a reader’s product expertise, three questions are asked. The respondents are asked to rate their knowledge of the products (a) in general, (b) compared to friends and acquaintances and (c) compared to experts (Brucks, 1985; Park et al, 1994). This was measured on a seven point likert scale (1= not knowledgeable, 7 = very knowledgeable). A total score for product expertise will be computed based on the average of these three items (Brucks, 1985; Park et al, 1994).
Next to these moderators, several control variables are included in the study. First, the respondents were asked to fill in their gender and age category. To measure the variable homophily, respondents evaluated choice-sets that where designed for their specific category and were not confronted with the other choice-sets. Due to constraints in the survey program, different routings in the survey were only possible based on one question. Therefore, the age and gender that were needed for these different routings were asked in a combined question. The age categories that were included are ‘younger than 20’, ’20-29’, ‘30-39’, ‘40-49’, ’50-59’ and ’60-69’ and ’70 or older’.
shopping statistics of the customers were retrieved. For each participating customer, the number of orders and the total amount of money they have spent in the last six months were recorded. Based on these two values, the average order value over the last six months could also be computed. Now that the moderators and control variables are discussed, the main attributes included in the Conjoint Analysis will be elaborated on.
Attributes and levels
The different readability and social factors that were hypothesized to influence review usefulness in the theory section, are included as attributes in the Choice-Based Conjoint Analysis. An overview of the attributes and levels is shown in table 1.
TABLE 1 Attributes and levels
To measure whether spelling errors influence the perceived usefulness of reviews, either no spelling errors are included in the review, or three. For degree of structure, three types of structure are chosen. The unstructured version is a text without spacing and headlines. The semi-structured review also starts with unstructured text, but concludes with the main advantages and disadvantages of the products. In the structured review, all arguments are shown under either the title ‘advantages’ or the title ‘disadvantages’. In order to measure real name exposure, three unisex names are chosen. For ‘nickname’, ‘first name’ and ‘full name’, these are respectively ‘Anoniem148’, ‘Anne’ and ‘Dominique de Winter’. This is done in order to prevent confusion because the variable gender changes between sets. For homophily, three demographics are used. Under the level ‘same’, all the demographics of the reviewer that are shown are equal to the respondent. For the level ‘different’, the opposite gender is used; Belgium is used as country (as only Dutch respondents will be included in the survey) and a category that is 20 years older or younger is used as different age. The different types of status a reviewer can receive are shown as labels under the name of the reviewer, whereas nothing is shown when the reviewer has no status. For the usefulness others attach to
Attribute Level 1 Level 2 Level 3 Level 4
Spelling errors No errors 3 errors
Degree of structure Unstructured Semi-structured Structured Real name exposure Nickname First name Full name Homophily between reader and
reviewer Different Same
Expert status of reviewer No status Top-reviewer Product-expert Usefulness others attach to the
review Negative votes (4) No votes Equal votes (2-2)
the review, four types of votes are included. On the bottom of the reviews either 4 negative votes; 0 votes; 2 positive and 2 negative votes or 4 positive votes are shown. As can be seen from table 1, not all the attributes have the same number of levels. A potential problem with this lies in the ‘number of levels effect’. The idea behind the number of levels effect is that the inclusion of an attribute with more levels than the others results in an increase in the
derived importance of that attribute relative to the other attributes (Steenkamp & Wittink,
1994). Although this potential problem is acknowledged, the choice is made to do vary the amount of levels among attributes in this study, in order to measure everything that was hypothesized. In designing the levels, it was at least aimed to not vary the amount of levels between attributes too much, which resulted in the levels that are shown in table 1.
Based on these attributes and levels, images that resemble reviews of both Dutch novels and tablets were created. In figure 2, example reviews from both categories are shown. These reviews were generated on a website that was set up for this research, since a lot of images needed to be created. After all versions were run, images were created by taking screenshots.
3.3. Study design
The designs of the choice-sets that were used in the surveys are created (with Sawtooth software SSI Web 6.6.6.) by means of complete enumeration. Complete enumeration generates designs that conform to a couple of principles. First, the designs have minimal overlap, which is desired since the probability that an attribute level occurs more than once in a single choice set should be minimized (Huber & Zwerina, 1996). Second, the levels should be balanced, which means that each level is shown approximately an equal number of times (Hair et al., 2010). Third, the design should be orthogonal, meaning that the attribute levels are chosen independently of the other attribute levels (Huber & Zwerina, 1996), so that the utilities can be measured independently.
Sawtooth software SSI WEB 6.6.6. provided the optimal design for the two versions within each product category. Based on the attributes and levels, the optimal design contains fourteen estimation tasks and one additional validation task. Thus, each version shows seven estimation sets and one validation set. Table 2 shows the efficiency rates of this choice task design. As can be seen, the minimum efficiency is 0.9620 for the levels ‘three errors’ within spelling errors and the level ‘same’ for homophily. Thus, despite the varying number of levels in this study, the design efficiency is pretty good.
TABLE 2
Efficiency rates of choice task design
Attribute Frequency Actual Ideal Efficiency
Spelling errors
No errors 21 Level deleted
3 errors 21 0,3337 0,3273 0,9620
Degree of structure
Unstructured 14 Level deleted
Semi-structured Structured 14 14 0,3817 0,3811 0,3780 0,3780 0,9803 0,9836 Real name exposure
Nickname 14 Level deleted
First name Full name 14 14 0,3841 0,3837 0,3780 0,3780 0,9685 0,9702 Homophily
Different 21 Level deleted
Same 21 0,3337 0,3273 0,9620
Expert status of reviewer
None 14 Level deleted
Top-reviewer Product-expert 14 14 0,3817 0,3811 0,3780 0,3780 0,9803 0,9836 Usefulness others attach to review
Thus, for each product category, 14 estimation sets and one validation set were shown. In each set, three different reviews are shown. To correctly measure the attribute homophily, different conjoint sets are created for different groups of consumers. The categories ‘younger than 20’ and ’70 or older’ were exposed to the same reviews as the categories ’20-29’ and ‘60-69’ respectively. This choice is made to reduce the loading time of the online survey and is at least partly justified by the fact that they are still equal on the other two demographics that measure homophily. Besides, from all categories, this age category is still closest to their own age. Thus, based on age and gender, 10 different groups that needed to follow a different routing in the survey were distinguished.
For both of the two product groups, 15 sets are made which all contain three different reviews. Figure 3 shows an example of a choice-set for Dutch novels. Since there were 15 different sets with each three different options, 45 pictured were needed for each of the ten categories of respondents. Thus, 450 pictures needed to be created for the ‘Dutch novel’-version, as well as for the ‘tablet’-novel’-version, resulting in 900 unique images.
A no-choice option was not included because of several reasons. First, consumers might simply choose this no-choice option to avoid difficult choices. Second, consumers can get bored at the end of the interview and decide to continue by choosing the no-choice option for the remaining questions, which leads to biased results. Furthermore, consumers might feel that certain attributes are missing, which can result in preferring the no-choice option (Gunasti & Ross, 2009). As it is essential for this research that consumers do make trade-offs, the no-choice option is not included.
The reviews are all modifications of real reviews from the website of the online retailer. In each set, the content of the reviews is the same, so that people cannot choose a review based on different preferences for product characteristics. Between the sets, however, different reviews are shown so that people will still perceive the reviews to be different. Also, it is tried to keep the variation in length and complexity of the reviews as low as possible. Each review has some negative points, but more positive points. To emphasise this, each review has an evaluation of four stars, so that there is no dispersion between reviews. The stars are included in the reviews because in practice, stars can be found in almost all reviews and hereby the reviews are a better representation of reality.
3.4. Plan of analysis
Construct validation
(KMO) should be higher than 0.5. If these statistics are sufficient, the rotated factor scores will be explored to see if variables have a loading above .40 on only one of the factors, and if the loadings are above .70 to see if more than one-half of the variance is accounted for by the loading on a single factor (Hair et al., 2006). If the underlying constructs can indeed be combined into a single factor the average value can be used as covariate in the CBC to measure whether the proposed moderators indeed affect the relationships between the attributes and review usefulness.
Independent Samples T-tests
Before the results can be discussed, there are two assumptions that need to be tested. First, it should be tested whether the products used in this study truly are medium- and high-involvement products. This was already confirmed by the pre-test, but that was only done with a limited sample (N=20). As a second validation, it will also be tested for the respondents of the main studies by means of an Independent Samples T-test. Furthermore, in the theory section, it was argued that reviews are more often used in the purchase of higher involvement products. Another independent T-test will be done in order to check whether that assumption actually holds.
In the Independent Samples T-tests, first the Levenes Test for Equality of Variances will be evaluated to check whether equal variances can be assumed. Then, the outcomes of the t-test will be interpreted to see if the means are significantly different.
Model specification and estimation
SEMI = Semi-structured NI = Nickname
FI = First name FU = Full name
H = Homophily between reader and reviewer NS = No status
TS = Top-reviewer status PS = Product expert status NEG = Negative votes NON = no votes EQ = equal votes POS = positive votes
In this model, the dependent variable is the chosen review that respondents select from the choice sets. The explanatory variables are the readability and social attributes.
Preference functions
Three preference functions will be considered for each attribute, namely a vector model (linear), ideal-point model (quadratic) and a part-worth function model. The part-worth function model estimates the largest amount of parameters, since separate estimates are made for each level (Green & Srinivasan, 1978). Linear models, on the other hand, use one specific parameter per attribute. Character variables are treated as nominal, whereas variables with numeric values are treated as linear. Also, attributes with only two levels are treated as linear. In this research, the remaining attributes with more than two levels are in essence character variables. The different part-worths for each attribute in the aggregated models will be plotted, to find out if some variables can be treated as numeric.
Model fit
Covariates
The moderators and the control variables will be included in the model as active covariates first, to see which of them are significant in explaining the effects. Then, the model will be estimated again with only the significant variables as active covariates and with the insignificant variables as inactive covariates. Also, the model will be estimated with all covariates as inactive covariates. From these models, the one with the best model fit will be chosen.
Validation
To see how well the final models predict, both the internal and external validity will be assessed. The internal validity will be analysed by means of the predicted cases computed by LatentGold, while the external validity will be tested by means of a holdout task. The holdout task for the Dutch-novel version is shown in figure 4. For the tablet-version, the attribute levels in the holdout task are identical.
In designing the holdout task, two things needed to be kept in mind. First of all, there is no point in asking people to choose among options where one dominates in such a way that everyone agrees it is best. Secondly, it should be avoided to design equally attractive options, since a completely random simulator would also predict equal shares. So, based on the suggestion of Orme (2010), it was tried to present concepts where the choices would lie somewhere around 50/30/20. In the validation set that is shown in figure 4, it was expected that option 1 would be preferred most often as five of the attributes levels are expected to have the highest utility, followed by option 3 where three options are expected to have a high utility. The least preferred option is expected to be option 2, where two of the attributes are expected to have a high utility.
For the external validity, the hit rate will be calculated. LatentGold will generate individual coefficients for each respondent. Based on these coefficients, the prediction probability will be calculated for each individual by dividing the exponent of the utility of each option by the sum of the exponents of the utilities for all three options. The option that is predicted is the option with the highest prediction probability. The predicted choices will then be compared to the actual choices they have made for the holdout set. Based on that information, the total hit rate can be computed.
4. Results
This section deals with the results from both the aggregated and segmented CBC analyses. Before these results are discussed, construct validity and two post-tests are discussed. Then, the outcomes of the CBC analysis for Dutch novels are discussed, followed by the outcomes of the CBC analysis for tablets. Since the hypotheses are tested for different product categories and on both aggregate and segment level, an overview of the hypothesis testing is given at the end of this section.
Sample characteristics
categories, there is not one single category that really stands out. What can be seen, however, is that for both versions the shares of the categories ‘<20’ and ‘69<’ are much smaller than the shares of the other age categories. Based on these demographic characteristics, it can be said that the respondents are a good representation of the entire customer base from the online retailer. The educational level for both samples does not show much difference, for both the category ‘HBO’ is largest, followed by the category ‘MBO’. The same goes for employment, where for both versions the largest category of consumers works ‘full-time’ and the smallest category consists out of ‘students’. The mean product expertise is almost equal for both versions (3,49 and 3,44 respectively). Although the amount of orders placed last half year does not differ much (2,91 and 2,97 respectively), the total amount spent and the average order value for the last six months are both higher for the consumers that participated in the ‘tablets’ version.
TABLE 3 Sample characteristics Covariates
Study 1 (medium-inv. products) (N=698)
Study 2 (high-inv. products) (N=744) # % # % Gender Male Female 274 424 39,3% 60,7% 317 427 42,6% 57,4% Age category < 20 20-29 30-39 40-49 50-59 60-69 69 < 18 92 121 156 167 113 31 2,6% 13,2% 17,3% 22,3% 23,9% 16,2% 4,4% 19 138 133 126 146 139 43 2,6% 18,5% 17,9% 16,9% 19,6% 18,7% 5,8% Education (highest)
Primary or secondary school MBO HBO University 56 224 279 139 8% 32,1% 40% 19,9% 70 241 296 137 9,4% 32,4% 39,8% 18,4% Employment Full-time Part-time Retired Student Currently unemployed 259 188 99 41 111 37,1% 26,9% 14,2% 5,9% 15,9% 275 171 128 78 92 37% 23% 17,1% 10,5% 12,4%
Mean St. Deviation Mean St. Deviation
Product expertise 3,49 1,34 3,44 1,46
Amount spent last 6 months 43,18 47,04 47,62 58,36
Number of orders last 6 months 2,91 2,74 2,97 3,86
Construct validity
The questions that consist of multiple scales need to be tested for construct validity. The convergent validity of all constructs is good, as the Cronbach’s alpha’s are all above 0,7 (PDI Dutch novel = 0,758, PDI tablets = 0,711, Dutch novel expertise = 0,894 and tablet expertise
= 0,905). The factor analysis shows that Bartlett’s test of spericity was significant (X2/df =
4680,527/66, p<0,01), indicating that at least some variables are correlated. The KMO measure of sampling adequacy is 0,687, which indicates that the pairs of variables that are correlated can be explained by other variables, since this value is higher than the minimum of 0.5. Since these statistics are both sufficient, the factor matrix will now be interpreted. As can be seen in table 4, all scores have a significant loading (>0,40) on only one factor. Moreover, all loadings are above 0,70, which means that more than half of the variance is accounted for by the loading of one factor (Hair et al., 2007).
TABLE 4 Rotated factor scores
Component PDI Dutch Novel PDI Tablet Expertise Dutch Novel Expertise Tablet PDI-DutchNovel1 0,867 PDI-DutchNovel2 0,882 PDI-DutchNovel3 0,714 PDI-Tablet1 0,886 PDI-Tablet2 0,851 PDI-Tablet3 0,701 Exp-DutchNovel1 0,952 Exp-DutchNovel2 0,937 Exp-DutchNovel3 0,896 Exp-Tablet1 0,952 Exp-Tablet2 0,949 Exp-Tablet3 0,903
Note: factor loadings less than .40 have been omitted from the table and variables have been sorted by loadings on each factor
Post-tests on importance of reviews and purchase decision involvement
significant (t(1356)= -15,388, p<0,01) and it can be concluded that the mean PDI for tablets (M=5,34, SD=1,12) is indeed higher than the mean PDI for Dutch novels (M=4,35, SD=1,30). Furthermore, in the theory section, it was argued that reviews were regarded as more important in the purchase of higher involvement products than in the purchase of medium involvement products. The Levene’s test of Equality of Means is again strongly significant (p<0,01), thus equal variances cannot be assumed here either. The T-test for unequal variances shows that this T-test is also strongly significant (t(1356)= -12,355, p<0,001). Thus, the answers from the survey confirm the assumption, as there is a significant difference between the extent to which consumers find reviews important when purchasing tablets (M=5,17, SD=1,21) compared to the importance of reviews when purchasing Dutch novels (M=4,28, SD=1,51). Thus, it seems to have been a good choice to use reviews from medium- and high-involvement products instead of low-involvement products, as the use of reviews is indeed regarded as more important for higher-involvement products. Another aspect that can be noticed here is that both means are above the average of the 7-point likert scale that was used, highlighting the importance of this research as reviews are thus of great importance for both medium and high-involvement products, albeit to different extents.
4.1. CBC Analysis for the medium-involvement product ‘Dutch novel’
Aggregated model
For each attribute, the estimated part-worths and relative importance are summarized in table 5. The relative importance is based on the range between the lowest and highest utility. The higher the range, the higher the relative importance is. Apart from real name exposure, all other attributes do have a significant impact on review usefulness. Degree of structure is seen as the most important attribute regarding review usefulness (60,49%). Overall, semi-structured reviews are considered to be most useful, followed by the semi-structured version. The second most important attribute is spelling errors (17,34%). Reviews without spelling errors are regarded as more useful than reviews that contain three spelling errors, and this effect is of second most importance in determining review usefulness. Reviews from reviewers that have more in common with the reader are indeed regarded as more useful, although the attribute is not very important overall (5,06%). Expert-status (10,29%) ranks third in importance. Reviews from reviewers with the label top-reviewer are regarded as more useful. The label product-expert does not have a significant effect on review usefulness. Furthermore, positive usefulness votes have a positive effect on review usefulness, whereas no votes make the review less useful. The relative importance of usefulness votes is not very high (6,82%).
Figure 5. Preference function plots -‐1
-‐0,5 0 0,5 1
unstructured semi-‐structured structured
Par ameter Level Structure -‐0,15 -‐0,1 -‐0,05 0 0,05 0,1 0,15
no status topreviewer product-‐expert
Par ameter Level Status -‐0,1 -‐0,05 0 0,05 0,1
negative none equal positive
Par
ameter
TABLE 5
Parameters and relative importance aggregated model
Model selection
For determining the optimal amount of segments, the model fit criteria are compared. First, all covariates are used to predict class membership. Table 6 shows the model fit criteria for models varying from one to six segments, where active covariates are used.
TABLE 6 Model selection
As can be seen from the table, a three-class model is preferred based on the CAIC and a four-class model is preferred based on the BIC. As a three-four-class model has more evenly divided group sizes, this model is preferred. Gender and education are found to be significant as active covariates (P<0,05). A 3-classmodel with gender and education as active covariates results in a better model fit (BIC = 8209,2237 & CAIC = 8246,2237) compared to a 3-class model where all covariates are active. However, a 3-class model where all covariates are inactive results in an even better model fit (BIC = 8186,8289 and CAIC = 8215,8289). Thus,
Attribute Parameters
Wald-statistic p-value Range
Relative importance Spelling errors -0,3686** 117,86 0,000 0,3686 17,34% Degree of structure Unstructured Semi-structured Structured -0,7295** 0,5566** 0,1728** 833,96 0,000 1,2861 60,49% Homophily 0,1076** 9,62 0,0019 0,1076 5,06%
Expert status of reviewer
None Top-reviewer Product-expert -0,0909** 0,1279** -0,037 38,50 0,000 0,2188 10,29%
Usefulness others attach
Negative votes (0-4) No votes (0-0) Equal votes (2-2) Positive votes (4-0) -0,0303 -0,0852** -0,0512 0,0642* 12,313 0,0064 0,145 6,82%
* = z-value > 1.96 and ** = z-value > 2.58
Model LL BIC (LL) CAIC (LL) df p-value R2
the covariates are only used to describe the segments. Because ‘real name exposure’ was also highly insignificant (p=0,52) on a three-class segment level, it was chosen to continue with the model without this variable because of the better model fit. This extra check was done in order to investigate the true structure of this model, because the estimates that are obtained from the aggregate model are typically biased when choices are based on different utilities for different segments (Orme, 2011). As the attribute is also highly insignificant for the three-class model, the finding that ‘real name exposure’ is not significant reflects the true structure of this model.
Segment interpretation
In table 7, the individual partworths for the three classes are shown, together with their relative importance for each attribute. As can be seen, the different classes clearly have different preferences. Based on these preferences, three different segments have been distinguished. Together with the values of the inactive covariates that are shown in table 8, a description of each segment is given.
TABLE 7
Parameters and relative importance on segment level
Note: per attribute, the segment with the highest relative importance is underlined
Attribute Segment 1 Size (0,412) Weight Segment 2 Size (0,341) Weight Segment 3 Size (0,247) Weight Spelling errors -0,7304** 11,89% -0,4249** 27,83% -0,6243** 13,13% Degree of structure Unstructured Semi-structured Structured -2,0343** 2,0288** 0,0056 66,12% 0,1133 -0,0096 -0,1036 14,20% -1,6140** 0,0257 1,5882** 67,36% Homophily -0,071 1,16% 0,2349** 15,38% 0,0834 1,75%
Expert status of reviewer
None Top-reviewer Product-expert -0,3233** 0,3613** -0,0380 11,14% -0,1316** 0,1184** 0,0132 16,37% -0,0980 0,3320** -0,2340** 11,91%
Usefulness others attach
Negative votes (0-4) No votes (0-0) Equal votes (2-2) Positive votes (4-0) 0,2236* -0,1476 0,2600** -0,3360** 9,70% -0,1442 -0,1294 0,0177 0,2559** 26,21% 0,1239 -0,1539 -0,0007 0,0307 5,84%
Segment 1 “Self-determined customers”
This segment consists out of more males than the other two segments (45,81%), although there is still a bigger share of females (54,19%). Compared to the other segments, people in this segment are relatively young. This segment contains the most full-time employed and the most students, compared to the other segments. From the highest educated people, the most fall in this segment. Furthermore, the total amount spent and the average order value in the last 6 months is highest for this segment.
For this segment, the degree of structure is most important (66,12%) and a semi-structured review is preferred most. Thus a u-inverted effect is present, as more structure is preferred to some extent, but a fully structured review is perceived less positive than a semi-structured review. Next, spelling errors (11,89%) and expert status (11,14%) are the most important for this segment. Spelling errors have a negative effect on review usefulness and reviews from reviewers with the label ‘top-reviewer’ are perceived to be more useful. What is striking for this segment, is that usefulness votes have a significant impact, but in the opposite direction from what was hypothesized. Negative votes have a positive impact whereas positive votes have a negative effect. Also, homophily is not very important for this segment (1,16%). Summarized, although they are sensitive for the top-reviewer label, it can be said that people in this segment neither care about what others think about the review, nor if the review is written by a similar individual. People in this segment are labeled ‘self-determined customers’.
Segment 2 “Confirmation-seekers”
In this segment, there are relatively more older people than in the other segments. From the lower-educated people, most fall in this segment (10,06% ‘primary or secondary school’, 37,82% ‘MBO’). This segment also contains the largest share of people that are retired (18,74%) or unemployed (17,16%). People in this segment have placed the highest number of orders in the last 6 months (2,95) but the amount they spent and the average order value is lowest of all three segments (121,97 and 38,62 respectively).