‘Words themselves are innocuous- it is the consensus that gives them true power.’

(1)

‘Words themselves are innocuous-

it is the consensus that gives

them true power.’

1

The Impact of Review Consensus and its Valance on the perceived

Usefulness of Online Consumer Reviews -

The Importance of Product Type, Product Involvement and Gender

Radostina Zlatanova June 27, 2016

(2)

The Impact of Review Consensus and its Valance on the perceived

Usefulness of Online Consumer Reviews -

The Importance of Product Type, Product Involvement and Gender

Master Thesis June 27, 2016

Radostina Zlatanova University of Groningen Faculty of Economics and Business

MSc Marketing

Specialization Marketing Management

Wielewaalplein 294 9713BR Groningen Tel: +31 (0)6 38496372 E-Mail: r.zlatanova@student.rug.nl

Student number: s3002314

(3)

Abstract

This study investigates the role of consensus (high vs. low) in a set of online consumer reviews and its valance (positive vs. negative) on consumers’ perception of usefulness of the same set and their purchase intentions. Additionally, the role the product type it terms of easy vs. difficult to evaluate was examined. The results show that high positive consensus (the majority of the reviewers agree on the positive performance of a product or service) is perceived as more useful compared to a low consensus set containing equal amount of positive and negative opinions. The effect of high positive consensus is stronger when consumers have high product involvement. In addition, the perceived usefulness of a set of online reviews is higher when consumers experi-ence differexperi-ences to judge the performance of the product. Lastly, the usefulness of a set of online reviews positively impacts consumers’ purchase intentions.

(4)

Acknowledgements

This thesis is submitted in fulfilment of the requirements for the Master's Degree in Marketing Management of University of Groningen. It contains work done from February to June 2016. Many people have contributed to the successful accomplishment of my master thesis whom I would like to show my gratitude for their support during this process. First, I would like to thank to my supervisor Dr. Liane Voerman for being always available and willing to answer all my questions, for her excellent guidance and constructive feedbacks during our meetings as well as for her support in moments of personal issues. I would like to thank to the other group members for their helpful participation during the meetings as well and to all participants, without whose cooperation the analysis would not have been possible. Subject related discussions with my par-ents Mila and Vladimir Zlatanovi, my sister Alexandrina Zlatanova and my grandfather Alexan-der Zlatanov as well as their wise advice, kind words and support were very heartening. A great note of thanks also to my friends Vanina Velikova, Denitsa Dimitrova, Annika Georgieva and Maya Spasova and to my boyfriend Sven Giesen for keeping me motivated and encouraging me mentally in times of downturns.

(5)

Making the perfect purchase decision nowadays is a very challenging task due to the great prod-uct offer offline as well as online, and the choice overload that consumers experience (Scheibe-henne et al., 2010). Therefore, when making a purchase decision, people often use different deci-sion aids in order to save time, money and efforts: they talk to experts like sales persons, ask friends and family members for advice, or use previous purchase experiences (Murray & Häubl, 2008). With the fast development of the Internet, customers can easily find many sources of in-formation about a certain offer online. In order to be better informed and to facilitate the decision making process, many customers use online reviews, comments, product experiences and rec-ommendations from other customers, i.e. the electronic word of mouth (eWOM) (Park & Lee, 2009). More precisely, eWOM refers to online word-of-mouth communication, which is wide-spread on various Internet platforms such as online forums, electronic bulletin board systems, blogs, review sites, and social networking sites (Goldsmith, 2006). A specific form of eWOM are the online consumer reviews (OCRs), defined as any positive or negative product or service in-formation provided by current or previous customers about their experiences, evaluations, and opinions (Park & Park, 2008).

Due to the growth of the internet usage, more and more customers engage in writing and reading online reviews (Lee et al., 2008). Online information, especially reviews and evaluation from previous customers, influences consumer attitudes no matter whether they purchase via Internet or at a traditional brick-and-mortar store (Muniz & O'Guinn, 2001). Research has shown that OCRs tend to be the second most trusted source of information after friends’ recommendations and to have a huge impact on consumer purchase decisions (Nielson, 2013). According to Niel-sen (2012), 70% of the consumers have a positive and trustful attitude towards online reviews from peers. Recognizing the crucial role of online reviews on firm success, many companies in-vest much in internet activities and eWOM management (Moorman, 2014). Due to the undoubted impact of this internet phenomenon on customer decision making process, the following study takes the OCRs as a main research objective.

1.1 The usefulness of OCRs

(8)

helpful-ness is defined as“ the extent to which consumers perceive the product review as being capable of facilitating judgment or purchase decisions” (Li et al., 2013). The OCR usefulness is an important driver for success for several reasons. Firstly, it has been shown that the OCR usefulness has a significant impact on consumers’ probability to adopt and use the information for their own pur-chase decisions (Cheung et al., 2008). Furthermore, Zhang et al. (2014) argue that reviews per-ceived as helpful drive customers’ purchase intentions as they provide important product related information. Hence, the perceived review usefulness has a significant impact in driving sales of online retailers (Hu et al., 2014). Realizing this, many online retailers like Amazon.com ask their customers “Was this review helpful for you?” in order to be able to show potential customers the most helpful reviews and accordingly boost sales. Therefore, the OCR perceived usefulness seems to be an important requirement for product attitude formation and buying intention. However, the majority of the scientific literature concentrates on finding out what makes a single review useful. In reality, when looking for additional product information online, customers face a multitude of OCRs, which contain positive as well as negative opinions for the same object and their purchase decision is influenced by a set of OCRs rather than a single review. Therefore, the following study aims to investigate the perceived usefulness of a set of OCRs by examining sev-eral factors influencing the usefulness perceptions of the customers.

1.2 The importance of OCRs consensus

(9)

Previous research has proven the important role of consensus information in influencing cus-tomer behavior. According to the attribution theory of Kelley (1967), when cuscus-tomers describe their response to a stimulus, i.e. their purchase decision, they are affected by the level of consen-sus among others. Hence, the level of agreement or disagreement among a group of people drives individual’s buying behavior. In addition, it has been shown that a positive eWOM message is more persuasive when the consensus across reviewers is higher and that the credibility of eWOM decreases when there is a high disagreement across the reviewers, meaning low consensus (Qiu et al., 2012). A question that arises is whether a positive consensus (most of the reviewers agree on the positive product performance) and a negative OCR consensus (most of the reviewers agree on the negative product performance), i.e. the valance of the consensus, have the same relative im-pact on customer responses.

Therefore, this study suggests that the level of agreement between two or more users regarding a product or its performance, called OCR consensus (Doh & Hwang, 2009) and its valence has a significant impact on the evaluation of a set of OCRs being useful or not.

1.3 The relevance of the product type

The importance of OCRs can be also explained by information economics. In many cases cus-tomers do not have perfect information about the product when making a purchase decision, in-cluding lack of information about the available alternatives, the product quality or supplier qual-ity. Due to this uncertainty, customers perceive a certain risk, which could take different forms, e.g. financial or performance risk (Urbany et al., 1989).

Yet, the level of perceived risk is not equal for all products and services. In order to decrease the perceived risk, customers search for additional product information, also labeled as search costs, acquiring not only physical but also cognitive processing efforts. However, as it is impossible to fully eliminate purchase uncertainty, consumers often compare the benefits of the additional search with its costs (Stigler, 1961).

(10)

assess the performance characteristics of the goods or not and whether this is possible before or

after the purchase, three different performance characteristics can be distinguished, which are

shown in Table 1.

Time of quality evaluation

before purchase after purchase

General possibility to assess the product

quality

possible search properties experience properties

impossible experience or

cre-dence properties credence properties Table 1: Product properties

Quality evaluation of a search good like a camera is possible prior to purchase while the quality of experience goods like a bottle of wine can be assessed after the product purchase and its usage or consumption. The product quality assessment of credence goods requires extremely high search costs and, therefore, is very difficult or even impossible even after purchasing (Darby & Karni, 1973). However, it is very important to point out that such a product classification strongly depends on the consumer’s individual and subjective perception and assessment (Wilde, 1981; Weiber & Adler, 1995). Additionally, the majority of the products consist of a mixture of search, experience and credence attributes. However, this categorization is still widely accepted (Huang et al., 2009).

(11)

1.4 The role of the product involvement

Another concept that plays an important role in the consumer decision making process is product involvement. Involvement with a product makes a consumer able and motivated to conduct prod-uct-related conversations with peers. Such an intense occupation with a product forms a lot of thoughts and emotions that can be easily recalled in WOM situations (Dichter, 1966). However, similar thoughts can be applied to the OCRs as a form of eWOM.

According to the elaboration likelihood model (ELM), consumers can process information in two ways depending on the level of involvement they have towards the purchase decision (Park & Kim, 2009). According to this study, customers who are motivated and able to process informa-tion, are more likely to engage in effortful processing of persuasive arguments, spend more time and thoughts before making the purchase decision. On the other hand, individuals lacking moti-vation or ability are more likely to process the information via peripheral routes by focusing on non-content cues. Therefore, the present study examines how the way consumers perceive certain information, i.e. the level of product involvement, impacts the usefulness perceptions in an OCR context.

1.5 Gender differences

In the 1990’s, 95% of the internet users were men (Weiser, 2000) with internet being perceived as a “boy toy” for tech-savvy users. However, the evolution of the internet to an everyday tool has led to extreme grow of female internet users (Weiser, 2000; Yang & Wu, 2006).

Several differences between males and females have been found regarding their online shopping behavior. Males show a more positive attitude towards purchasing product and services online and are more likely than females to engage in online shopping (Rodgers & Harris, 2003; Slyke et al., 2002). Additionally, the same authors find that men tend to trust online shopping more and are more satisfied with its outcomes than women and, therefore, use it more often. This is due to the fact, that females perceived risk with regard to online shopping is higher as well as their ap-prehension when providing private information like phone numbers, credit card numbers, and addresses online (Bartel-Sheehan, 1999).

(12)

standing. Furthermore, product recommendations from friends have a greater impact on females than on males affecting the risk perceptions and willingness to purchase the product (Garbarino & Strahilevitz, 2004). Therefore, one can assume that women are more likely to rely on OCRs in purchase decisions. In summary, gender is an interesting variable to look at when investigating the impact of OCR consensus on OCR usefulness.

1.6 The effect on purchase intention

Due to the spread of commercial web sites and the increased acceptance of online transactions by consumers, the electronic commerce is developing very fast (Hong et al., 2004). The online shopping environment differs from the traditional brick and mortar stores in several manners (Al-ba et al., 1997). One of the main differences is that while in offline shops consumers can touch or smell the products, in online shopping environment they have to base their purchase judgements on presented information on the website. In order to overcome this limitation, many online retail-ers allow their customretail-ers to share their product experience and evaluation on the website, which facilitates the decision making process (Chatterjee, 2001). As stated earlier, previous research shows that the review helpfulness has a positive impact on purchase intention, meaning that when customers perceive a certain review as helpful, their purchase intention goes up. This study will also examine whether this finding will hold for a set of OCRs.

1.7 Problem statement and research questions

The main objective of this paper is to investigate the impact of consensus and the valance of this consensus in a set of reviews on the perceptions of the customers whether this set is useful or not. In order to investigate the topic accurately, the study focuses on the following research questions:

 How does the OCR consensus impact OCRs usefulness perceptions?

 How does the valance of the OCR consensus (positive vs. negative) influence these per-ceptions?

 How does the product type influence the evaluation of usefulness of an OCR set?  How does the level of involvement influence the usefulness evaluation of an OCR set?  Does the gender moderate the relationship between review consensus and review

(13)

 Overall research question:

How do the OCR consensus and its valence influence the perceived OCR usefulness and behaviour intentions depending on the product type, involvement level and the gender?

1.8 Structure of the thesis

(14)

2 Theoretical framework

The theoretical framework gives additional background information and explanation of relevant theories in current academic literature about the determinants mentioned in the introduction. Firstly, the independent variables in this research are presented and explained, including the OCR consensus and the product type and their effect on perceived OCRs usefulness. After that, the interaction between the independent variables is explained. Lastly, the moderating roles of prod-uct involvement and gender are elaborated. An overview of all hypotheses and the conceptual framework finalize this chapter.

2.1 Purchase Intention

Previous research has found out that information usefulness has a strong positive impact on in-formation adoption (Cheung et al., 2008). Additionally, Zhang et al. (2014) found out that the source quality positively affects purchase intention. This finding was supported by Park et al. (2007).

These findings refer to single reviews. However, the same thoughts can be applied to a set of OCRs. Therefore, this study suggests that high perceived usefulness of a set of OCRs will lead to high purchase intention:

H1: The higher the perceived OCR usefulness, the higher the purchase intention. 2.2 The effect of OCRs consensus

(15)

opin-ions from different reviewers are present, the higher the strength of consensus (Burnkrant & Cousineau, 1975; Kelley, 1967). Therefore, in the OCRs context, if a person sees that two or more people agree on the performance level of a product, he will feel more confident that this performance is true.

Usually, consumers tend to look for others’ opinion when they want to decrease the cognitive effort and uncertainty related to a purchase decision (Dowling & Staelin, 1994). However, re-search has shown that the uncertainty increases when there is a low or no consensus among oth-ers’ viewpoints which leads to a negative reaction according to such uncertainty, meaning a rejec-tion of the viewpoint (West & Broniarczyk, 1998). Hence, a high proporrejec-tion of agreeing opinions assure the consumer that the information is trustful and acceptable (Chiou & Cheng, 2003). In summary, this study suggests that consensus information about a product on the Internet in form of unbalanced set of OCRs is more persuasive and trustworthy than conflicting information about the same product or balanced set of OCRs. Hence, the following is expected:

H2a: The higher the review consensus or unbalance (both positive and negative) among the reviewers within a set of OCRs, the higher the perceived OCR usefulness of the same set.

2.3 The effect of the valance of the OCRs consensus

Review writers can agree on either a good product performance or a bad product performance, indicating that OCR consensus can be either positive or negative. Prior research has shown that the diagnostic value of a single review might depend on whether it is a positive or negative in-formation (Bone, 1995) due to the fact that consumers perceive negative inin-formation as more useful and diagnostic compared to positive information when forming attitudes and overall evaluations towards a subject, referring to the so called negativity effect (Skowronski & Carlston, 1989). According to these theories, one of the main factors affecting the informative value of a cue is its discrepancy from expectation. Furthermore, due to the fact that people expect most life outcomes to be positive and positive information is closer to the normative expectancy than is negative information, negative information should get more weight than positive one when cues are combined into impression (Skowronski & Carlston, 1989).

(16)

positive (or neutral) information about products or services as it seems to be less useful in catego-rizing them, because it is more ambiguous and any product or service can have one or more posi-tive aspects (Ahluwalia et al., 2000; Herr et al., 1991). In summary, an object is more easily as-signed to a low quality category due to negative information, than to a high quality category based on positive information. Therefore, negative information can be assumed as more useful and diagnostic when making an evaluation decision.

These theoretical hypotheses have been empirically proven. Previous research has shown that negative eWOM has a greater impact on brand evaluation compared to positive eWOM (Chiou & Cheng, 2003). Additionally, Sen and Lerman (2007) argue that consumers are more likely to pay attention to negative rather than to positive online reviews, concluding that consumers perceive negative product evaluation as more useful than a positive one. The negativity effect implies that negative cues appear more attractive than positive ones and that customer pay more attention to negative information, which can be explained by the fact that negative information has more di-agnostic value (Kanouse & Hanson, 1972; Sen & Lerman, 2007). Hence:

H2b: Negative opinion consensus has a stronger positive impact on the perceived OCR useful-ness compared to positive opinion consensus.

2.4 The effect of the product type

Consumers may not only seek for others’ opinion in order to reduce their cognitive effort and uncertainty, but also because there is no information about the product or because existing infor-mation is not informative enough (Bone, 1995). The author states that the attribute inforinfor-mation is often hard to check especially for experience attributes of a product such as the drivability of a car.

(17)

a bottle of red wine. The value of credence goods can never really be known with certainty, like car repair, medical treatment, and education (Hsieh et al., 2005).

The prior scientific research has paid far more attention on search and experience goods com-pared to credence goods (Tsao, 2014; Lim & Chung, 2011; Park & Lee, 2008). However, this study suggests that consumers tend to rely extremely on OCRs before purchasing a credence good. This is firstly due to the fact that the information asymmetry between seller and buyer of credence goods is larger compared to search and experience goods (Hsieh et al., 2005). As a re-sult, customers need more information about credence goods and are compelled to rely on exter-nal cues such as OCRs. The information need combined with the unwillingness of the most cus-tomers to share their opinions with regard to credence goods, results in high diagnosticity and value of OCRs about this type of goods (Feldman & Lynch, 1988; Pan & Chiou, 2011). Thus, credence goods are expected to be related to higher perceived risk as the assessment of these goods is challenging even after their purchase and use and online information provided by co-customers should be extremely useful (Mitra et al., 1999).

Due to the intangible nature of the credence goods, clear and determinable product attributes and benefits are missing. Therefore, online information from other customers is likely to be very im-portant and convincing for consumers willing to purchase credence goods (King & Balasubrama-nian, 1994; Zeithaml, 1981). Thus, consumers tend to put more effort in searching online infor-mation about credence goods and assess it carefully. As consumers usually do not have other information sources, online reviews are perceived to be professional, probative and diagnostic, which leads to credibility of the information and its further adoption (Feldman & Lynch, 1988; Lynch et al., 1988).

(18)

compared to experience and search goods. Consequently, it is expected that the usefulness of online reviews depends on the product nature. According, the following hypothesis is proposed:

H3a: The more difficult it is for the customer to evaluate a certain product or service, the higher the perceived OCR usefulness.

2.5 The interaction between consensus and product type

As discussed above, in situation in which consumers cannot evaluate the product quality due to vagueness in judging criteria, external information that facilitates their purchase decision is highly valuable (Bone, 1995). Consistent with the objective of risk reduction, this paper suggests that the higher the uncertainty and the lesser product information available, the more useful and helpful the online review. For consumers willing to buy credence goods, product related informa-tion tends to be scarce and difficult to process. It would become even more difficult for the con-sumers to process the information when there are conflicting viewpoints. This means that a bal-anced set (so, no consensus) of OCRs related to products with most credence properties, includ-ing equal proportion of positive and negative will confuse the customer and hinder him from the purchase. In contrast, when purchasing search goods consumers are more likely to believe in their own judgments which are based on external commercial information making them more passive in their acceptance of eWOM (Pan & Chiou, 2011). Therefore, this study suggests that the OCR consensus is of a great importance for customers intending to purchase a credence good and of a lower relevance for customers intending to purchase experience goods and even lower for search goods. Thus, the following hypothesis is proposed:

H3b: The positive impact of OCR consensus on perceived usefulness is more pronounced for products and services which are difficult to evaluate like credence goods compared to products and services that are easier to evaluate like experience and search goods.

2.6 The moderating effect of the product involvement

(19)

1988). In other words, the likelihood that customers elaborate information processing depends on the level of involvement they show towards the object.

The relationship between involvement and information processing can be explained with the elaboration likelihood model which refers to the phenomenon that customers can process the same information in two different ways depending on their involvement (Petty & Cacioppo, 1986). OCRs typically have two functions: as informant, providing user-oriented product infor-mation and as recommender, providing inforinfor-mation about personal product experiences from previous customers. According to the ELM, these types of information can be processed either through the central or the peripheral route. ELM postulates that those consumers who are able and motivated to process a message tend to do that via the central route (Maclnnis et al., 1991). They typically engage in a thoughtful, effortful, and enduring processing of persuasive argu-ments, devote themselves to these arguargu-ments, and at the end of the process create their own thoughts with regard to the arguments. On the other hand, people who have neither the motiva-tion nor the ability or lack both are more likely to process the informamotiva-tion via peripheral routes or mental shortcuts by focusing on non-content cues (Petty & Cacioppo, 1984).

(20)

re-view quality. They are motivated to extract as more product information as possible in order to have sufficient reasons for purchasing the product. In this case, a more controversial set of online reviews may appear more interesting and useful for them, as it will present both up- and down-sides of the product.

In summary, this study suggests that the level of involvement plays a moderating role when evaluating a set of online reviews. Under low-involvement conditions, consumers tend to rely on peripheral cues, like number of positive of negative OCRs, and not on product-related informa-tion. On the other hand, in high involvement situations customers are motivated to engage in ef-fortful processes of information comprehension and elaboration. Hence, the study hypothesizes the following:

H4: The positive impact of consensus information within a set of OCRs on perceived OCR use-fulness tends to be higher when product involvement is low.

2.7 Gender differences

Prior research shows that there are significant differences between males and females in the rea-sons why both genders engage in WOM communication and how they perceive it. The females’ greater desire for social connection and their higher tendency to rely on and be open to peers’ opinions results in more active WOM and eWOM engagement (Bae & Lee, 2011; Eagly & Carli, 1981; Kempf & Palan, 2006).

This significant difference in being influenced by other consumers’ opinions when making a pur-chase decision can be explained by the selectivity theory, which states that females tend to proc-ess information in amore effortful, extensive and detailed way while males tend to rely on heuris-tics (Meyers-Levy, 1988). In addition, the difference between females’ socialization desire and males’ independence desire can explain the gender difference in terms of reliance on peers’ rec-ommendations (Bybee & Zigler, 1990).

(21)

H5a: A set of OCRs is perceived as more useful by females than by males. 2.8 The interaction between consensus and gender

Furthermore, it is interesting and relevant to prove whether females react differently than males when evaluating a consensus set of OCRs. Darley and Smith (1995) report important gender dif-ferences in personality. The authors argue that males tend to be confident, independent, competi-tive, willing to take risks and accordingly less vulnerable to perceived risk compared to females. According to the selectivity theory, while women consider and evaluate all available information they receive, men use a selectivity strategy in order to save time and efforts. Moreover, men make decisions more quickly, focusing on objective cues such as physical attributes and relying on their own judgment (Meyers-Levy & Sternthal, 1991). In contrast, women tend to process information in a more detailed and comprehensive way, considering multiple external informa-tion sources rather than own judgements. Addiinforma-tionally, Todorov et al. (2002) show that inconsis-tent information force individuals to perform systematic processing. Hence, in a situation with a high degree of inconsistent information, customers may feel that a higher level of elaboration on such information is required to make judgments (Davis & Tuttle, 2013). Thus, females are ex-pected to be more responsive to an inconsistent set of OCRs and to perceive it as more useful than male consumers. Hence, the following is hypothesized:

H5b: The positive impact of consensus information within a set of OCRs on perceived OCR usefulness tends to be higher for males than for females.

2.9 Conceptual model and hypotheses overview

Table 2 gives on overview of the aforementioned hypotheses which will be tested in chapter 4.

H1: The higher the perceived OCR usefulness, the higher the purchase intention.

H2a:

The higher the review consensus or unbalance (both positive and negative) among the reviewers within a set of OCRs, the higher the perceived OCR usefulness of the same set.

H2b: Negative opinion consensus has a stronger positive impact on the perceived OCR use-fulness compared to positive opinion consensus.

(22)

H3b:

The positive impact of OCR consensus on perceived usefulness is more pronounced for products and services which are difficult to evaluate like credence goods compared to products and services that are easier to evaluate like experience and search goods. H4: The positive impact of consensus information within a set of OCRs on perceived OCR

usefulness tends to be higher when product involvement is low. H5a: A set of OCRs is perceived as more useful by females than by males.

H5b: The positive impact of consensus information within a set of OCRs on perceived OCR usefulness tends to be higher for females than for males.

Table 2: Hypotheses overview

The conceptual framework presented on Figure 1 serves to illustrate the proposed relationships in this study, testing the effects of the OCR consensus, its valance and the product type on the per-ceptions of OCR usefulness and buying intention depending on the product involvement and the gender of the consumers.

(23)

3 Research Design

This study aims to investigate the effect of consensus of the information within a set of online consumer reviews (in terms of proportion between positive and negative reviews) on the ceived usefulness of the same set. In addition, the product type is expected to influence the per-ceived usefulness as well as to interact with the OCR consensus. Lastly, individuals’ gender and involvement are expected to moderate the relationship between OCR consensus and perceived usefulness.

3.1 Study design

This study uses a 3 (OCR consensus: high positive, high negative and low consensus) x 2 (prod-uct type: search to experience and experience to credence) between subject factorial design to test the proposed hypotheses. Each subject was randomly assigned to one of the six experiment con-ditions shown in Table 3. According to Hair et al. (2009) in order to be reliable, the study needs at least 30 respondents per condition and, therefore, 180 respondents in total, being family mem-bers, friends, fellow students or third party respondents.

OCR Consensus Product type

High positive OCR consensus

Low OCR consensus (both positive and

negative)

High negative OCR consensus Search to experience Condition 1 Condition 2 Condition 3 Experience to

cre-dence Condition 4 Condition 5 Condition 6

Table 3: Overview of the 3x2 factorial design

(24)

peculi-arities of the particular machine”. The specialized knowledge required to provide services with credence qualities makes it more difficult for the consumer to evaluate the performance and qual-ity even after purchase and use. Therefore, the following research chooses a vacation in a hotel as a search to experience good, and a laptop repair service as an experience to credence good.

Figure 2: Product/ Service Evaluation Continuum 3.2 Sample

The sample of 205 respondents (43% males and 57% females) was randomly assigned to the six different scenarios. The half of the respondents was provided with scenario, asking them to imag-ine that they want to go on a vacation and see an onlimag-ine advisement followed by eight recent online reviews. The other half of the sample was presented to a scenario, telling them to imagine that their laptop needs a repair and they see an advertisement of a computer repair shop supported also by eight recent online reviews. The age of the respondents ranged between 16 and 64 years with an average of 28 years. The majority of the participants (66.3%) were between 21 and 27. Most of the respondents were either students (46.3%) or full-time employed (42.8%). The distri-bution in the different conditions was similar to these overall demographic results.

3.3 Survey development

The participants in the study were exposed to one of the six different sets of OCRs. The reviews for the three conditions for a vacation in hotel “del Rey” in Spain were collected from the real Web site Booking.com and these ones for the three conditions for a laptop repair shop "Computer Repair- any brand, any problem" from Google Reviews. The names of the hotel and the laptop repair shop exist in the reality. All reviews were modified in order to be appropriate for the study. The review length was kept constant as it could affect the reader’s judgement (Chevalier and Mayzlin, 2006). The reviews were created in a similar way to Google Reviews with the intention to make them more familiar to the respondents and more similar to a real life situation. Colorful bubbles with a capital letter indicate the name of the reviewer. The reviews were accompanied Easy to

evaluate

Difficult to evaluate

High in credence qualities High in search qualities High in experience qualities

(25)

with star rating in order to make it easier for the readers to remember the opinion of the reviewer. Additionally, an overall star rating was placed above the set of OCRs similar to the product re-view part of Amazon.com to give the reader an overre-view of the overall opinion distribution. The sequence of positive and negative reviews was the same in each condition in order to ensure that the results would not be confounded. Regarding the content of the reviews, the OCRs about vaca-tion consist of three components: locavaca-tion, food and staff. The OCRs about the laptop repair ser-vice also consist of three components: price, time spent for the repair and staff. Figure 3 shows one of the six conditions.

(26)

On Booking.com and other websites usually ten reviews appear on one page. However, in order to keep the respondents motivated to read all reviews, the amount of OCRs per set was reduced to eight. In order to test the external validity, at the end of the survey the participants were asked whether they usually read OCRs and if yes, how many they read. The results show that usually the majority of the respondents “definitely” (50.2%) or “probably” (33.3%) read OCRs. A bigger part of the participants (60.3%) indicated that they usually read either 6 to 9 or more than 9 OCRs which supports the choice of eight reviews per set.

The survey was translated in Bulgarian language as a considerable part of the planned partici-pants are Bulgarians. The survey was distributed online via e-mails and on social media platforms containing a link to the survey to respondents with different nationalities including Dutch, Ger-man, Bulgarian, Swiss, and Chinese.

3.4 Measures and Manipulations

Most of the measures used in the survey are based on existing literature. Some of them are a bit modified in order to better fit the designed concept.

Usefulness. After presenting one of the six possible scenarios to the respondents, they were asked

to rate the usefulness of the OCR set based on a 7-point Likert scale introduced by Bailey and Pearson (1983) with a Cronbach’s Alpha of 0.902.

Purchase Intention. The other dependent variable in the model, purchase intention, was

meas-ured on a 7-point Likert scale, containing four items proposed by Coyle and Thorson (2001) with an internal consistency of 0.83.

Consensus. The OCR consensus was manipulated by varying the proportion of positive and

(27)

Product type. The product type was also manipulated by assigning the respondents to either a

vacation in a hotel (experience good) or to a laptop repair shop (credence good). Following the procedure proposed by Krishnan and Hartline (2001) the respondents were asked to report their ability to judge the service/product performance before (first scale) and after (second scale) pur-chase and use on a 7-point scale (1=totally agree and 7=totally disagree).

Based on these questions, goods and services having a low score on both scales are interpreted as search goods as their performance can be assessed even before the purchase. Accordingly, goods and services with a high score on the first, but low on the second item are defined as experience goods as consumers are not able to judge the product performance until purchasing it. Lastly, goods and services having high spores on both scales are classified as credence goods because customers cannot evaluate the product performance even after purchase and use.

Perceived Risk. In addition to this, a second manipulation check for the product type was

in-cluded based on the risk, consumers perceive with the purchase with respect to the two prod-ucts/services. The 7-point semantic differential scale contains four items, proposed by Jain and Srinivasan (1990) with Cronbach’s Alpha of 0.80.

Product Involvement. In the scientific literature there are several different involvement

meas-urement scales. However, the most popular and used scale is the one proposed by Zaichkowsky (1985). It includes twenty semantic differential items scored on 7-point scales applicable to ad-vertisements, products and purchase decisions. However, a scale including twenty items is pretty extensive. Therefore, this research would use the modified version of this scale invented by Mit-tal (1995). The modified five-item personal involvement inventory (PII) showed an adequate evi-dence of unidimensionality and internal consistency as a measure for both purchase decision in-volvement and personal inin-volvement. The reliability of the modified construct as well as the cap-tured variance were very high, 0.90 and 0.64 respectively. Based on that and on the additional clear and easy structure, this study uses this modified involvement scale by Mittal (1995).

OCR Attitude. The attitude towards online reviews in general was added as a control variable.

(28)

Table 4 represents the scales used in the survey, the items of which they consist and the internal consistency measured by the authors.

Concept Items Source

Consensus

1. Most of the online reviews in the set shared the same

opinion.

2. There were many controversial opinions among the reviews in the set.

3. What was the overall reviewer opinion?

-

Product type

1. The service/product performance is easy to be evalu-ated before purchase and use.

2. The service/product performance is easy to be evalu-ated after purchase and use.

Krishnan and Hartline (2001)

Perceived Risk

1. It is really annoying to make an unsuitable choice - It is not annoying to make an unsuitable choice

2. There is little to lose by choosing poorly - There is a lot to lose by choosing poorly

3. I am certain of my choice - I am uncertain of my choice

4. A poor choice would be upsetting - A poor choice wouldn't be upsetting

Jain and Srinivasan (1990) Cronbach’s α=0.80

Usefulness

1. I found the set of online reviews useful.

2. The set of online reviews helped me to shape my atti-tude towards a vacation in hotel “del Rey”/the repair service of the computer repair shop.

3. The set of online reviews helped me to make a decision regarding a vacation in hotel “del Rey”/the repair ser-vice of the computer repair shop.

Bailey and Pearson (1983) Cronbach’s α=0.902

Purchase Intention

1. It is likely that I will book hotel "del Rey “. /It is likely that I will use the service of "Computer Repair- any brand, any problem".

2. I will book hotel "del Rey" the next time I want to go on a vacation. /I will go to "Computer Repair- any brand, any problem“ the next time I need a laptop re-pair.

3. Suppose that a friend of mine calls me to get my advice in his search for a vacation, I will recommend him to book hotel "del Rey". /Suppose that a friend of mine calls me to get my advice in his search for a laptop re-pair shop, I will recommend him to go to "Computer Repair- any brand, any problem".

4. I will definitely visit hotel "del Rey". /I will definitely try the service of "Computer Repair- any brand, any problem".

(29)

Involve-ment

When going on a vacation, the hotel I am going to stay in is:/ When my laptop is broken and needs a repair, the re-pair shop I bring my laptop to, is:

1. Important – unimportant

2. Means a lot to me – means nothing to me 3. Matters to me – does not matter to me 4. Significant – insignificant 5. Of no concern – of concern to me Mittal (1995) Cronbach’s α=0.90 OCR Attitude

1. Before booking a certain hotel for a vacation/choosing a certain laptop repair shop, I always look for online reviews from other customers.

2. When booking a hotel for a vacation/choosing a certain laptop repair shop, online reviews from peers are help-ful for my decision.

3. Online customers’ reviews do not make me feel confi-dent in booking a hotel for a vacation.

Park, Lee & Han (2007) Cronbach’s α=0.79

Table 4: Measurement scales 3.5 Preparation of the data: Validity and Reliability

As a first step, the scales used for measuring the different concepts were tested for their validity and reliability. Adjusting of the scales was performed if needed. Next, it has to be proven, whether the manipulation checks worked out appropriately. For this purpose, one-way analysis of variance (ANOVA) as wells as Paired Sample Test were performed. Thereafter, in order to measure the effects of the independent variables and their interactions on the dependent variables, a multiple linear regression is performed.

Originally the data set consisted of 249 responses. However, 41 of the respondents were excluded as they did not fill in the survey appropriately, meaning that they either did not answered the questions measuring the dependent variables or answered many questions automatically without reading them. In order to prepare the data set for further analysis, the verification of the scales construction through factor analysis was executed.

(30)

sig-nificant, meaning that the H0=items are uncorrelated can be rejected. Furthermore, the commu-nalities should exceed 0.5.

Additionally, in order to measure the internal consistency a reliability analysis was performed. Scales with a Cronbach’s Alpha greater than 0.6 were determined as reliable (Malhorta, 2009 p.319). Additionally, it was proven whether the Cronbach’s Alpha increases if items are deleted. As a first step, an overall factor analysis including all items was conducted in order to indicate the separate concepts. As a result, almost all concepts were derived correctly according to the predetermined scales with communalities higher than 0.5 and without cross loadings, except for the perceived risk scale. Thereafter, an individual factor analysis for each concept was run. Table 5 shows the criteria for the factor and reliability analyses and their results. The strike through items indicate that they are removed from the scales based on the results of the reliability analy-sis. FA Criteria Perceived Risk Usefulness Purchase Intention Involve-ment OCR Attitude KMO> 0.6 0.533 0.727 0.873 0.861 0.590

Bartlett’s Test of

Spheric-ity 0.000 0.000 0.000 0.000 0.000 Variance explained 67.03% 77.99% 90.77% 69.83% 60.77% Communalities > 0.5 1. 0.669 2. 0.590 3. 0.719 4. 0.703 1. 0.748 2. 0.816 3. 0.776 1. 0.907 2. 0.914 3. 0.895 4. 0.913 1. 0.746 2. 0.841 3. 0.833 4. 0.827 5. 0.244 1. 0.736 2. 0.726 3. 0.361 2 Factors: 1+4 & 2+3 Eigenvalues 1. 1.559 2. 1.123 2.340 3.631 3.492 1.523 Cronbach’s Alpha > 0.6 0.156 0.858 0.97 0.642 0.626

Cronbach’s Alpha if item

deleted 0.616 - - 0.930 0.961

Table 5: Factor and Reliability analysis

Perceived Risk. The perceived risk is measured with four items, explaining 67.03% of the

(31)

low (0.156), indicating internal inconsistency. As it can be seen in Table 5, items 1 and 4 load highly on one factor and items 2 and 3 on a second factor. A possible explanation for this could be that item 2 is worded negatively, which was perceived by the participants differently from the other items. Moreover, the reliability analysis shows that if items 2 and 3 are deleted, the Cron-bach’s Alpha and, therefore, the internal consistency will increase to 0.61. Additionally, putting items 2 and 3 into one factor, results in Cronbach’s Alpha of 0.248. Based on these results, item 2 and item 3 are removed from the perceived risk scale.

Usefulness. The four items used for measuring the perceived usefulness, explain 77.99% of the

variance. The KMO value (0.727) is higher than required. The Bartlett’s test of Sphericity is sig-nificant (p<0.001) for all constructs. The Cronbach’s Alpha is 0.858, indicating that the scale is consistent and reliable. Deleting an item does not lead to higher internal consistency.

Purchase Intention. The purchase intention scale perfectly met all requirements with a KMO

value of 0.873, significant Bartlett’s Test of Sphericity, 90.77% explained variance, communal-ities higher than 0.8 and Cronbach’s Alpha of 0.97. Deleting an item only leads to a decrease in the internal consistency.

Involvement. Looking at the involvement concept, the items explain 69.73% of the variance in

involvement and the sample adequacy is more than sufficient with value of a 0.861. Four out of five items’ communalities are higher than 0.7. The last item “of concern to me” has a communal-ity of 0.244 and, therefore, lower than the required 0.5 level. Additionally, by removing this item the internal consistency increases to 0.93. Based on these findings, item 5 is removed from the involvement scale.

OCR Attitude. The OCR attitude was measured with three items, explaining 60.77% of the

(32)

3.6 Manipulation Checks

In order to prove whether the participants perceived the difference between the conditions in terms of consensus and product type as such, the manipulations were checked using one-way ANOVA.

3.6.1 Consensus

The manipulation checks about consensus included two questions asking the respondents whether the set of the reviews included similar or controversial opinions. Additionally, the respondents were asked a third question about the overall opinion of the hotel, or the laptop repair shop re-spectively, testing whether the respondents noticed the valance of the consensus.

Firstly, a one-way ANOVA is performed proving whether there is a significant difference be-tween the means of the high and low consensus conditions. For this purpose, the highly positive and the highly negative consensus conditions are combined into one group, called “high consen-sus”. The results in Table 6.1, indicate that the mean of the high consensus condition is signifi-cantly higher (4.92) compared to the mean of the low consensus (3.67) regarding the question whether the OCR set included similar opinions. Moreover, regarding the second question, whether there were many controversial opinions, the results for high consensus conditions are significantly lower (4.56) compared to the low consensus (5.62) case. However, the mean of the high consensus is relatively high. This could be due to the fact that in the case of the high consen-sus sets, the proportion was 6:2 and not 10:0.

Most of the online re-views in the set shared

the same opinion.

There were many

controver-sial opinions among the

re-views in the set.

Consensus _{μ (SD)} _μ(SD)

1. High consensus 4.92*** (1.67) 4.56*** (1.70) 2. Low consensus 3.67*** (1.90) 5.62*** (1.23) * p < 0.1, ** p < 0.05, *** p <0.00

Table 6.1: Manipulation check results for consensus

Additionally, a second ANOVA is performed, splitting the high consensus condition into highly

positive and highly negative consensus and comparing them to the low consensus condition. The

(33)

is no significant difference between the highly positive and highly negative consensus conditions (p=0.19), but that there is a significant difference between the high positive consensus and low consensus set (p<0.001) and between the high negative consensus and the low consensus scenario (p=0.004). The respondents again correctly indicate the difference between the two high consen-sus conditions (positive and negative) and the low consenconsen-sus condition (5.19 and 4.66 > 3.67 on “Most of the online reviews in the set shared the same opinion.” & 4.52 and 4.60 < 5.62 on “There were many controversial opinions among the reviews in the set.”).

The Turkey post-hoc test on the third question (“What was the overall reviewer opinion?”) re-vealed that there is a significant difference between all three consensus conditions (p<0.03). As shown in Table 6.2, the respondents correctly recognized the consensus valance rating the overall reviewer opinion of the high positive consensus condition in average with 5.29, of the low consensus set with equal amount of both positive and negative reviews with 4.76 and of the high negative consensus set with 3.10.

Table 6.2: Manipulation check results for consensus and its valence 3.6.2 Product type

In the next step, it was proved whether the two products representing different product types, namely mostly experience and mostly credence, were perceived as different. For this purpose, the respondents should indicate on a 7-point Likert scale from 1=totally disagree to 7=totally agree whether it is easy to judge the quality and performance of the product or service firstly before and secondly after its purchase and use. The results show that there is no significant difference (p=0.146) between the means on the first item, indicating that both services were difficult to judge before purchase and use. However, there is a significant difference (p < 0.001) between the two services in terms of judging the performance after use. The respondents assigned to the vaca-tion condivaca-tion found it more easy to judge its performance after stay in the hotel compared to the respondents assigned to the laptop repair shop condition judging the performance of the repair shop after use of the service (6.02 > 4.72). In order to test whether the means of the two related

What was the overall

(34)

groups “before” and “after” are significantly different from each other, a Paired Samples Test is performed. The results demonstrate that the manipulation check worked out well indicating a significant difference (p<0.001) between the two groups. Table 7 represents the means and the standard deviation of the different product types.

The quality is easy to be evaluated BEFORE stay/use

The quality is easy to be evaluated AFTER stay/use

μ (SD) μ (SD)

Vacation in a hotel 4.43 (1.41) 6.02*** (1.02)

Laptop repair 4.08 (1.80) 4.72*** (1.27)

Total 4.25 (1.63) 5.35*** (1.32)

* p < 0.1, ** p < 0.05, *** p <0.001

Table 7: Manipulation check results for product type

In addition to this, a second manipulation check for the product type was included based on the risk, consumers perceive with the purchase with respect to the two products/services. According to the literature, the perceived risk should be higher for products and services that are more diffi-cult to evaluate. Therefore, another ANOVA was run in order to test this assumption. The results do not indicate a significant difference (p-value=0.259) between the means of the vacation (4.72) and the laptop repair scenario (4.59). However, based on the results of the first manipulation check presented in Table 7, the manipulation of the product type is accepted as successful.

3.7 Plan of analysis

In order to measure the effects of the independent variables and their interactions on the depend-ent variables, a multiple linear regression is performed, using the following formula:

Usefulness of OCRs/ Purchase Intention= β0 + β1 × High Positive Consensus + β2 × High

Negative Consensus + β3 × Product Type + β4× (High Positive Consensus × Product Type) +

β5 × (High Negative Consensus × Product Type)+ β6 × Involvement + β7× (Involvement × High

Positive Consensus) + β8× (Involvement × High Negative Consensus) + β9 × Gender + β10 ×

(High Positive Consensus × Gender) + β11× (High Negative Consensus × Gender) + OCR

Atti-tude + OCR Behavior + Age + Income + ε

(35)

consensus is transformed into two dummy variables (0=low consensus and 1=high posi-tive/negative consensus). A dummy variable was created also for the product type variable with 0 indicating an experience good and 1- a credence good. In the second model, the interaction be-tween consensus and product type is added. Model 3 is extended with the involvement and its interaction with consensus. The forth model includes the gender variable and its interaction with consensus. In the last model, all independent variables, the moderator, their interactions as well as the control variables are included.

It is important to mention that each variable that was measured with a 7-point scale was mean centered. The main reason for mean centering the scales is yielding a proper interpretation of the interaction effects in the regression analysis (Aiken & West, 1991). Mean centering is a method by which the mean of a certain variable becomes a value of zero. This leads to change in the in-tercept, but not in the regression coefficients for the mean centered variables. In an uncentered regression model, the intercept is interpreted as the value for the dependent variable when any-thing else is zero. However, such an interpretation would not make sense here as the scales used in this research ranged from 1 to 7 and value of 0 would indicate very useless OCR set, very un-involved costumer and so on.

Interestingly, the involvement and the perceived OCR usefulness have relatively high mean val-ues of 5.5 for involvement and 5.1 for usefulness, indicating that in the given sample the average consumer was rather highly involved and found the OCR set rather useful according to the given 7-point Likert scales with an average of 3.5.

It was also suggested that when consumers find a set of OCR useful, it is more likely that they will purchase the product. Therefore, the purchase intention variable was added into the model as a second dependent variable. It is supposed that the OCR consensus and the product type have impact on the customer’s behavioral intention, but that this effect is mediated by the OCR per-ceived usefulness. In order to test this assumption, a mediation analysis was run in SPSS using the PROCESS macro by Hayes (2013). According to Baron and Kenny (1996), some require-ments have to be fulfilled when assessing the mediation effect:

(36)

 The mediation variable should have a significant effect on the dependable variable and  Including the mediation variable in the model, the independent variable should not have

an effect on the dependent variable (full mediation) or at least substantially lower effect on the dependent variable (partial mediation)

Figure 4 shows a typical mediation model.

X = causal variable Y = outcome variable M = Mediator c = total effect

c´= direct effect of X on Y

ab = indirect effect of X on Y (the product of a and b)

c = c' + ab

(37)

4 Results

In this chapter, the proposed hypotheses will be tested by means of a multiple linear regression analysis. Firstly, a two-way ANOVA is run in order to prove whether there is a significant differ-ence between the means of the two groups (vacation in a hotel vs. laptop repair) on the OCR use-fulness as a dependent variable and whether there is an interaction between the two main inde-pendent variables with OCR usefulness (see Table 8). The results show that the two main effects are significant but that there is no significant interaction between the consensus level and the product type, as the p-value for the interaction term is 0.848. The descriptive statistics table shows that, independent from the product type, the high positive consensus condition is rated as most useful, followed by the high negative consensus set. As at least useful the participants rated the low consensus condition. Additionally, the total OCR usefulness of the three consensus con-ditions is 0.3406 higher for the laptop repair shop compared to the total average score of the va-cation cases.

A) Vacation in a

hotel B) Laptop repair Total significant _different from μ (SD) μ (SD) μ (SD) 1.High positive consensus 5.0103 (1.2690) 5.8385 (0.7029) 5.4308 (1.0960) 2** 2.Low Consensus 4.5429 (1.6029) 5.1350 (1.2236) 4.8257 (1.4546) 1** 3.High negative consensus 4.9169 (1.5977) 5.5311 (1.1877) 5.2462 (1.4163) Not Significant Total _{4.8148 (1.5002)} _{5.5063 (1.0949)} _{5.1657 (1.3522)} significant different from

B) Laptop repair*** A) Vacation in a hotel*** * p < 0.1, ** p < 0.05, *** p <0.001

Table 8: ANOVA results for the OCR usefulness 4.1 Assumptions for regression analysis

(38)

that there should be no multicollinearity, meaning that the predictor variables are not highly cor-related with each other, and no heteroscedasticity, meaning that all observations have equal vari-ances in the disturbance term. Furthermore, the assumption of a normal distribution of the error term should also be met.

Multicollinearity. As a first step it should be proven whether multicollinearity is an issue or

whether here is a relationship pattern between the prediction variables. If this is the case, this would lead to unreliable parameter estimates. In order to detect multicollinearity, a correlation matrix of the predictor variables is computed to prove whether they are highly correlated mean-ing correlations of above 0.80 and 0.90 (Field, 2009 p.325). The correlation matrix (see Appendix A) does not show high significant correlations between the independent variables. The high posi-tive consensus and the high negaposi-tive conditions are correlated with each other (Pearson Correla-tion= -0.500, p<0.01) which is not surprising as they both represent the high consensus condition and gender is correlated with positive consensus (Pearson Correlation= -0.152, p=0.034). At this stage, the results do not indicate multicollinearity. However, this assumption will be proved again when estimating the regression through the values of the variance inflation factor (VIF), which should not exceed 10 and VIF values > 4 are worthy of concern (Field, 2009 p.325).

Heteroscedasticity. The next assumption supposes that the error term is homoscedastic, meaning

that it has the same variance in all cases across different sections. Therefore, it will be tested whether the disturbance terms have equal variance across the two different product types and the three different consensus conditions by applying the Levene’s test. Table 9 shows the results of the Levene’s test, indicating that the error terms are homogenetic, as the test is insignificant for both product type and consensus condition.

Normality. Lastly, the normality assumption should be tested. Nonmorality may lead to changes

in the estimates. In case of nonnormality, the p-values of the parameters are not meaningful as they are linked to the normality assumption. In order to test this assumption, the

Kolmogorov-LeveneStatistic df1 df2 Significance

Product type _1.449 ₁ ₁₉₃ _0.230

Consensus

Condition 2.384 2 192 0.095

(39)

Smirnov and Shapiro-Wilk tests are performed. Both tests are not significant (p Kolmogorov-Smirnov= 0.2 and pShapiro-Wilk= 0.745), indicating that error term is distributed normally.

4.2 Estimation

After checking all assumptions, the multiple regression model can be estimated. The full model looks like as follows:

Usefulness of OCRs/ Purchase Intention= β0 + β1 × High Positive Consensus + β2 × High

Negative Consensus + β3 × Product Type + β4× (High Positive Consensus × Product Type) +

β5 × (High Negative Consensus × Product Type)+ β6 × Involvement + β7× (Involvement × High

Positive Consensus) + β8× (Involvement × High Negative Consensus) + β9 × Gender + β10 ×

(High Positive Consensus × Gender) + β11× (High Negative Consensus × Gender) + OCR

Atti-tude + OCR Behavior + Age + Income + ε

The OCR consensus and the product type were included as multiple dummy variables, with a low consensus and a vacation in a hotel respectively as base lines.

As a first step, a regression analysis is performed on the base model of this research, Model 1, including the main independent variables, testing their effects on the OCR usefulness as a de-pendent variable. Model 2 includes also the interaction effect between the two indede-pendent vari-ables. The third model includes the involvement effect and its interaction with consensus. In model 4, gender and its interaction with consensus are integrated. Model 5 represents the full model with all independent variables, moderators and control variables, and their interactions. However, as we will see later on, adding the interaction terms of gender and consensus leads to VIF scores higher than 10, indicating multicollinearity. Therefore, Model 6 represents the final model excluding the interactions of gender with consensus. All results are presented in Table 11.

Variable Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

‘Words themselves are innocuous- it is the consensus that gives them true power.’