INITIAL TRUST IN WEBSHOP RECOMMENDATIONS
A Choice Based Conjoint on Online Recommender Agents
INITIAL TRUST IN WEBSHOP RECOMMENDATIONS
A Choice Based Conjoint on Online Recommender Agents
Master thesis, Marketing Management and Marketing Intelligence
University of Groningen, Faculty of Economics and Business
3
MANAGEMENT SUMMARY
The aim of this research is to investigate to what extent explanations, familiar products and information about the product, in the output of an Online Recommender Agent (RA), have an impact on a users’ initial trust in the RA. In this context the degree of product expertise possessed by the user is also taken into account.
Choice Based Conjoint Analysis was used to obtain the responses from a sample of 202 people. Respondents were asked to choose the most trustworthy RA from three hypothetical RAs, which differed on the levels of four attributes namely explanations, number of familiar products included, detailed product information and links to independent reviews. The last two attributes are categorized as ‘information about a product’. The attribute ‘Explanations’ includes the levels ‘how’, ‘why’ and ‘trade-off’ explanations. The attribute ‘inclusion of familiar products’ has four levels, increasing from 0-4 familiar products. Detailed product information includes the levels ‘picture’ and ‘text’, while links to independent reviews includes the levels ‘link to expert reviews’ and ‘link to consumer reviews’. Every attribute has a ‘no option’ to be able to measure the effect on initial trust relative to not including any of the levels. Two questionnaires, with 9 choice sets each, were randomly distributed.
The results show that both the how and a trade-off explanation have a positive effect on a users’ initial trust, while the for the why explanation no effect on a users’ initial trust is found. When comparing these explanations, the trade-off explanation has the strongest effect on a users’ initial trust. For the inclusion of familiar products a positive linear relation is found, meaning that adding a familiar product in an RA, will increase the initial trust of a user in an RA. Text, picture, link to consumer reviews and link to expert reviews also show a positive effect on a users’ initial trust in an RA.
4
With the results of this paper managers and RA developers can design an RA that maximizes the initial trust of its users. More specifically it enables managers to make trade-offs based on the characteristics of the customers, namely the level of product expertise, in order to maximize the initial trust of the users in the RA.
Keywords: Recommender Agent, Choice Based Conjoint Analysis, initial trust, familiarity,
5 TABLE OF CONTENTS MANAGEMENT SUMMARY 3 1. INTRODUCTION 7 2. THEORY 11 2.1Initial trust 11 2.2Explanations 12 2.3Familiarity 15 2.4Product information 17 2.5Product expertise 18 3. CONCEPTUAL MODEL 20 4. METHODOLOGY 21 4.1.Questionnaire design 21 4.2.Simuli 21 4.2.1 How explanation 22 4.2.2 Why explanation 22 4.2.3 Trade-off explanation 23
4.2.4 Familiar products in the recommendation 23 4.2.5 Detailed product information 24 4.2.6 Links to independent reviews 24
4.3 Choice Based Conjoint design 24
4.4 Product expertise 26
4.5 Comfort with computers and online shopping 26 4.6 Demographics and psychographics 26
4.7 Familiarity pre test 27
4.8 Data collection 27
4.9 Data analysis 28
4.9.1 Factor analysis 28
4.9.2 Familiarity check 28
6
5. RESULTS 31
5.1Descriptive analysis 31
5.2Choice Based Conjoint analysis 32
5.2.1 Aggregate results 32
5.2.2 Model fit aggregate model 32 5.2.3 Hypothesis testing aggregate results 33 5.2.4 Selection number of segments 34
5.2.5 Results segment analysis 36
5.2.6 Hypothesis testing on segment level 39
5.2.7 Predictive validity 40
6. DISCUSSION 41
6.1Conclusion 41
6.2Managerial implications 43
6.3Limitations and future research 44
REFERENCES 47
APPENDICES 57
Appendix I: Questionnaire 57
Appendix II: Sawtooth tests 70
Appendix III: LSD Test of the pre test 70
Appendix IV: Factor analysis 74
7
1. INTRODUCTION
The internet has opened a large number of opportunities for both businesses and consumers. The benefits for consumers are researched extensively, among them the most popular are transparency of prices, wider choice of products and lower search costs (Bakos, 1997; Diehl, Kornish, & Lynch Jr, 2003). Online retailers benefit by acquiring information from customers that can help them customize products and services. Industry experts indicate that personalizing shopping experience is an effective way for online retailers to turn browser into buyers (Verton, 2001). Recommendation agents are used as a tool to personalize the shopping experience (Häubl & Trifts, 2000). Recommender agents (RAs) are personalized computer agents that provide a consumer online recommendations on what product to buy, based on what an individual consumer needs (Komiak & Benbasat, 2006). Considering the overwhelming amount of information a consumer encounters in an e-commerce environment and, in most cases, the absence of real-time human assistance in online shops, RAs have become an essential tool for making the online purchasing process effective and efficient (Adomavicius et al., 2005; Ahn, 2006; H. Wang & Doong, 2010; Xiao & Benbasat, 2007).
The design of an RA is comprised of three components, namely the input, the process and the output (Xiao & Benbasat, 2007). In the input component user preferences are implicitly or explicitly elicited, the process component is where the recommendations are generated, while in the output the recommendations are presented to the user. Since the entrance of RA, research has focused mainly on process; the development of the algorithms used to generate recommendations (Adomavicius et al., 2005; Herlocker, Konstan, Terveen, & Riedl, 2004; Middleton, Shadbolt & De Roure, 2004) and thus failing to focus on both input and output. This therefore is a research that focuses on the output of an RA, more specifically how the output, or content of a recommendation, influences the perceptions of initial trust of a user in an RA.
8
unfamiliar recommendations have been found to lower users’ evaluations of a music CD RA (Cooke, Sujan, Sujan, & Weitz, 2002). Explanations of, for example, why and how an RA made recommendations could give insight in the process, or reasoning, of the RA. Following that reasoning Ye & Johnson (1995) found that explanation facilities increase the acceptance of and confidence in recommendations of expert systems. Providing information of the recommended product can make the recommendation more attractive to the user. In line with that reasoning Cooke et al. (2002) found that additional information of a new product could increase the attractiveness of an unfamiliar recommendation.
Trust is important in situations in which two parties experience dependence and where this dependence entails risk (Luo, 2002). Due to the large amount of information and complexity when searching in the e-commerce environment, customers depend on RAs for better decision making. Also risk arises because the quality of the RA, or its reasoning, is uncertain. Thus RA adoption will to a large extent depend on the trust of the user in the RA (Komiak & Benbasat, 2006). Researchers have noted the importance of initial trust, especially for new technologies (Harrison McKnight, Choudhury, & Kacmar, 2002). Also as trust in different stages exists out of different factors (McKnight, Cummings, & Chervany, 1998) a detailed investigation of initial trust is relevant. This study therefore focuses on initial trust, and more specifically initial trusting beliefs.
9
verify the abilities and skills of the RA (Hollander & Rassuli, 1999). Because of both the possible goal incongruence and information asymmetry a user might have concerns whether the RA acts in its interests and/or has the ability to do so, which can result in agency problems (Bhattacherjee, 1998).
These aspects of the relationship between the user and the RA hamper trust building. Therefore the focus of this research is to investigate whether trust between the RA and user can be improved, more specifically:
To what extent do explanations, familiar products and information about the product in the output of a Recommender Agent have an impact on a users’ initial trust in the Recommender Agent?
Trust in an e-commerce environment can be influenced by the expertise of the consumer. Gefen, Karahanna, & Straub (2003) have found that, while using a website, experts rely more on perceived usefulness when making transaction decisions, while novices rely more on trust. This indicates that, at least in the e-commerce environment, the importance of trust decreases with experience (Gefen, Benbasat, & Pavlou, 2008) Therefore in addition to the research question stated above:
Does the effect of the output of a Recommender Agent on a users’ initial trust differ for product experts versus product novices?
This research is conducted by using choice-based conjoint analysis, by which the contribution of each selected dimension of the output of an RA on intitial trust can be investigated on both aggregate and segment level. Initial trust will be used as dependent variable in this research. To create a deeper understanding tough, the underlying structures and reasons of initial trust, trusting beliefs, are discussed in the theory section.
10
Previous research has found effects on trust of personalization and familiarity in the process of the RA (Komiak & Benbasat, 2006) and explanation facilities in the input part of the RA (W. Wang & Benbasat, 2007). These researches are mainly focused on the input and process part of an RA though. To the best of my knowledge, specifically the output of an RA with respect to trust has not been investigated yet. Following that reasoning the effects of inclusion of familiar products, information about the product and explanations and their relative importance in that context has not been investigated also. According to W. Wang & Benbasat (2008) understanding trust formation in RAs will assist both researchers and practitioners in recognizing the right designs that infer trust with users, which will increase the effectiveness and acceptance of RAs. Enhancing the trust in technologies, will enable organizations to attract more customers and will help build relationships with customers(Reibstein, 2002).
11
2. THEORY
2.1 Initial Trust
Across disciplines trust is seen as important in situations of interdependence between two parties, in which risk exists (Rousseau, Sitkin, Burt, & Camerer, 1998). Trust makes understanding less complex by subjectively taking away the possibility that the trustee will behave undesirable (Gefen, Karahanna, & Straub, 2003). In the context of the relationship between users and RA’s, users depend on recommender systems for better decision making, while risk arises because there is a possibility the information provided by the RAs is of uncertain quality. Thus relying on an RA could make the users vulnerable to wrong decisions (Komiak & Benbasat, 2006). Therefore in order for RAs to be effective, users must have trust in the RAs’ product recommendations (Häubl & Murray, 2001).
In this research initial trust is investigated, which is trust formation in an early stage during which users interact with an unfamiliar party (McKnight, Choudhury, & Kacmar, 2002). Initial trust is relevant, because perceptions of uncertainty and risk are especially salient with the interaction of a user with an unfamiliar RA (W. Wang & Benbasat, 2007). According to McKnight et al. (1998) initial trust refers to trust with someone which is unfamiliar to the trustor. It is a relationship in which the actors do not yet have meaningful or credible information about each other nor have created an affective relationship with each other.
12
trustor that the trustee has the ability, skills and expertise to do what the trustor needs. Benevolence is the confidence of the trustor that the trustee cares about the trustor and has the motivation to act in the trustor’s interest. While integrity is the confidence of the trustor that the trustee is honest and keeps its promises, thus abide to a set of principles that is accepted by the trustor.
The notion of trust above refers to interpersonal trust, were the trustee is a human being. This research extends interpersonal trust to trust in a technology, were the trustee is a recommender agent, as also discussed in recent research (Corritore, Kracher, & Wiedenbeck, 2003; Li, Hess, & Valacich, 2008). Personalities are assigned to computers by people and although the personalities, or human properties, do not exist in technology, they are perceived to be so by users in their interaction with computers (Moon & Nass, 1996). The relationship between technology and humans has previously been discussed in the context of trust and technology acceptance (Komiak & Benbasat, 2006).
During a first interaction with an RA, a user creates trusting beliefs based on initial trust because there is a lack of previous interactions. In this stage the user will look for any cues or information in order to evaluate the RA’s trustworthiness (Gefen, Karahanna, & Straub, 2003). The output of a recommender system can infer those cues and therefore, in the following sections different possibilities that influence initial trust will be discussed.
2.2 Explanations
Explanations ‘serve to clarify and make something understandable’ (Gregor & Benbasat, 1999, p.498). Explanations can be perceived from two different aspects (Gregor & Benbasat, 1999). Firstly, the explanation can be commenced by the provider of information, in order to clarify, justify or convince. Secondly it can be commenced by an information receiver, to resolve disagreement of misunderstanding.
13
Explanation facilities have been researched in two focus areas, Knowledge Based System (KBS) explanations and Decision Support Systems (DSS) decisional guidance. While KBS research focused most on explanation facilities that give insight into creating transparency, thus why and how certain processes are used (Dhaliwal & Benbasat, 1996; Gregor & Benbasat, 1999). DSS research focused more on explanation facilities aiding users in their decision (Barkhi, 2002; Silver, 1991) W. Wang & Benbasat (2007) therefore used three different types of explanations in relation to the RA namely why explanations, how explanations and trade off explanations, an approach adapted in this study.
How explanations are used to show the user the line of reasoning used by the RA. The explanations give insight into the processes used to generate the recommendation (W. Wang & Benbasat, 2007). An RA is ought to produce correct and relevant recommendations based on the preferences and needs of the user, which most commonly is done by either content-based or collaborative-based filtering (Ansari et al., 2000). Content-based filtering generates recommendations based upon attributes that are desired by the consumer, while collaborative-based filtering uses the needs of like-minded consumers to generate recommendations (Ansari et al., 2000). When an RA does not explain how it produced recommendations it will create a knowledge gap between the RA and the user, since the user lacks knowledge of the process of the RA. In that case users will not be able to verify the RA’s ability and expertise (Hollander & Rassuli, 1999). This is consistent with the competence belief, since the confidence that the user has in abilities of the RA will be low (McKnight & Chervany, 2002). Previous research also showed that using technologies might have a negative influence on users’ trust, especially when information asymmetries occur (Hampton-Sosa & Koufaris, 2005). A how explanation will bridge the knowledge gap by giving insight in the process and thereby solving the information asymmetry. Therefore it is proposed that:
H1: Users will have more initial trust in RAs with how explanations than in those without how explanations.
14
in intention is consistent with the goal incongruence found in the agency theory (Keil et al., 2000). It also relates to the benevolence belief as the confidence of the user in that the RA has the motivation to act in the users’ best interest will be low. It is advantageous for an RA provider to use why explanations, because transparent capabilities are desirable and signal to users that the RA is the type they are searching (Bergen et al., 1992). Also since RA’s are internet based, fewer cues are present for users to create perceptions, compared to real sales persons of whom the intentions can be diverted from attitude, tone, appearance and so on (W. Wang & Benbasat, 2007). Previous research that showed that trust is created when someone understands another person’s goals and intentions better (Doney & Cannon, 1997). A why explanation will bridge the gap of intention by showing it acts in the interest of the user and thereby solving the goal incongruence. Therefore it is proposed that:
H2: Users will have more initial trust in the RAs with why explanations than in those without why explanations.
15
H3: Users will have more initial trust in the RAs with trade-off explanations than in those without trade-off explanations.
Differences between these explanations can be made by using Toulmin’s model of argumentation (Gregor & Benbasat, 1999). This is a model of human reasoning that has been used as ground for constructing explaining capabilities and provides a basis for the assessment of practical reasoning and argumentation (Toulmin, Rieke, & Janik, 1984). Since the model has the elements that are present in convincing arguments, explanations that conform more to the model lead to greater trust, satisfaction and agreement (Gregor & Benbasat, 1999). The model exists out of six elements, namely (1) claims, or the conclusions that are presented (2) grounds, or the factual information that is the basis of the argument (3) warrants, or the justification for going from the grounds to the claims (4) backing, or the authorization for the warrant (5) qualifiers, or statements about the degree of certainty of a claim and (6) possible rebuttals, or extraordinary circumstances that might interrupt the strength of the argument.
In an RA situation, the claims are the recommendations presented. The why explanation conforms to the model by providing a ground, through linking certain preferences of the user with the recommendations. The how explanation gives a form of warrant, by explaining the process of the RA, but by doing that does not provide the ground itself. The trade-off explanation on the other hand conforms to the model of Toulmin in that it both states possible
rebuttals, in the form of liabilities of a product, but also implicitly states a ground. Trade-off
explanations namely provide positive characteristics of each product, which could serve as ground for the respondents of why a certain recommendation is made. Trade-off explanations thus conform most to the model of Toulmin. Therefore it is proposed that:
H4: The initial trust of users in an RA will be more influenced by trade off explanations than by how and why explanations.
2.3 Familiarity of the recommendation
16
able to predict the behaviour and with that avoid surprises (Xiao & Benbasat, 2007). ‘Familiarity is an individual’s understanding of an entity, in many cases based on previous experience, interactions and learning of ‘the what, who, how and when of what is happening’ (Gefen, Karahanna, & Straub, 2003, p. 63). Therefore familiarity builds trust (Komiak & Benbasat, 2006).
Trust can flow from one entity to another (Doney & Cannon, 1997; Stewart, 2003). W. Wang & Benbasat (2008) indicate that when users interact with an RA for the first time, they identify heuristics and entities closely linked to the RA, in order to assess its trustworthiness. Following that reasoning, when a trustworthy source, closely associated with the RA, is present, it will increase trust in the RA (W. Wang & Benbasat, 2008). The products in the recommendations are closely linked to the RA, and thus when familiar, can infer trust.
Behavioural research suggest that consumers view unfamiliar recommendations negatively (Park & Lessig, 1981). Only presenting unfamiliar recommendations, therefore, might give the user the impression the RA does not have the intention to fulfill its needs, something an RA is ought to do. This will thus create an intention gap. This intention gap is consistent with the goal incongruence found in the agency theory (Keil et al., 2000).
When evaluating an RA, a user focuses on recommendations that are diagnostic to the abilities of the agent (Cooke et al., 2002). The unfamiliar recommendations are less diagnostic than familiar recommendations by definition (Cooke et al., 2002), but familiar recommendations will not infer trust by definition. In order to infer trust the user has to be confident that the RA has the ability to do what the user needs, consistent with the competence belief (McKnight & Chervany, 2002).
A negative perception about a product will be an indication of an unwanted product, therefore such a ‘negative’ familiar recommendation will lower the confidence of the user that the RA has the ability to do what the user needs. Conversely familiar recommendations related to positive perceptions will infer trust, and can solve the intention gap. Therefore it is proposed that:
17 2.4 Product information
In e-commerce, the main interaction of the e-vendor with the consumers, is through a web site. In order for the e-vendor to communicate commitment to a consumer, an obvious way is through the character of the web site. Gefen, Karahanna, & Straub (2003) state that when more effort is placed in creating a web site that is usable, consumers will find the web site is easy to use and conclude that the e-vendor is investing in the relationship. This notion can be extended to RAs, as they are used in the same context. Thus a usable interface of an RA, can signal that the e-vendor of the RA invests effort in creating a relationship between the user and the e-e-vendor, and thus that he cares about the user. Since trust may come from heuristic cues like an impression (Jarvenpaa, Shaw & Staples, 2004; McKnight et al., 1998) and heuristic cues can create trust quickly, even before having a firsthand experience (Jarvenpaa, Shaw & Staples, 2004), an interface can have an effect on initial trust.
On the contrary a unusable interface could create negative cognitions leading to a reduction of trust (W. Wang & Benbasat, 2008). The negative cognitions thus indicate the opposite of what is described above, namely that an RA does not care for the user. This is consistent with the problem of goal incongruence found in the agency theory (Keil et al., 2000).
Detailed product information in the form of product descriptions in text or multimedia can signal to the users that the RA acts in their interest and cares about them(Xiao & Benbasat, 2007). The detailed product information will namely make the RA more useful, since it will create a more detailed insight of the product. By doing that the detailed information will thus contribute to the user’s assessment of the benevolence of the RA (Xiao & Benbasat, 2007). This means that the provision of detailed information of a product in an RA will create trust with the user and will solve the goal incongruence. Therefore it is proposed that:
H6: Users will have more initial trust in RAs with detailed information of the product than in those without detailed information of the product.
18
consumers can verify if the recommendations provided by the RA are desirable and whether the claims of the RA are true (W. Wang & Benbasat, 2008). Presenting links to such independent reviews will create a more useful interface, as the users can gain more of such independent information. It will also create a sense among the users that an RA behaves both in an unbiased fashion and cares about the users. This will contribute to the assessment of the integrity and benevolence beliefs of the users (McKnight & Chervany, 2002). Thus including independent reviews in an RA will create trust with the user and will solve the goal incongruence. Therefore it is proposed that:
H7: Users will have more initial trust in RAs with links to independent reviews than in those without links to independent reviews.
2.5 Product expertise
19
extended to the RA-user relationship. Why explanations, detailed information of the product and links to independent reviews have been identified as having an effect on trust by indicating the e-vendor cares about the user thus through effecting the benevolence belief. Therefore it is proposed that the initial trust of users with low product expertise will affected more by why explanations, detailed product information and independent links compared to users with high product expertise. On the other hand the ‘how explanation’ and ‘inclusion of familiar products’ are related to the competence belief, while the ‘trade off explanation’ to the integrity belief. Since these beliefs are not in line with the notion of benevolence, these variables are expected not to be moderated by product expertise. This therefore results in the following hypotheses:
H8: The effect of why explanations on the initial trust of users in RAs will be larger for users with low product expertise than users with high product expertise
H9: The effect of detailed information of the product on the initial trust of users in RAs will be larger for users with low product expertise than users with high product expertise.
20
3. CONCEPTUAL MODEL
Figure 1: Conceptual model
Consumer heterogeneity
‘Market segmentation involves viewing a heterogeneous market as a number of smaller homogeneous markets, in response to differing preferences, attributable to the desires of customers for more precise satisfactions of their varying wants’ (Smith, 1956, p.6). Market segmentation presupposes heterogeneity in the preferences of buyers for products or services (Green & Krieger, 1991). As market segmentation is widely used by companies in order to position their products or services effectively (Green & Krieger, 1991), consumer heterogeneity is relevant to investigate. The heterogeneity in the preferences for products or services can be related to person variables like demographic characteristics, psychographic characteristics and product or service usage (Green & DeSarbo, 1979) therefore several of these variables are taken in account in the research approach.
Explanations: - How explanation - Why explanation - Trade-off explanation Product expertise Detailed product information: - Pictures - Text Inclusion of familiar products
21
4. METHODOLOGY
This research uses Choice Based Conjoint (CBC) to identify the effect, and relative importance of each of the attributes on initial trust. The goal is to gather at 200 respondents, since a sample size of 200, in usual applications of CBC, will provide an acceptable margin of error (Hair et al., 2010). These respondents will have to have at least purchased online once, to ensure the questions are interpreted correctly.
4.1 Questionnaire design
The questionnaire (Appendix I) is created in Qualtrics and build up the following way. Firstly a check is done whether the smartphones used in the survey are perceived correctly, in the sense of familiarity and goodwill. The reason for the placement of these questions in the beginning is that familiarity can be falsely inferred after taking the survey (doing the Choice Based Conjoint). After these questions the respondents will be presented with nine Choice Based Conjoint questions, each consisting of three profiles. This will be followed by four questions, aimed to identify respondents with product expertise. The questionnaire ends with questions regarding the usage situation of an RA and questions regarding demographic and psychographic characteristics. The specifics of these questions will be discussed in the rest of this section.
4.2 Stimuli
22 Attributes Levels Explanations - No explanation - How explanation - Why explanation - Trade-off explanation
Familiar products - 0 familiar products - 1 familiar product - 2 familiar product - 3 familiar products
Detailed product information - No detailed information - Text
- Picture
Links to independent reviews - No link to independent reviews - Link to expert review
- Link to other consumers’ evaluation
Table 1: Attributes and levels
4.2.1 How explanation. The how explanation is used to show the user the line of reasoning used by the RA, which is based on their preferences and needs. In practise this comes down to a short sentence indicating how the recommendations are made, thus which process is used. A technique used a lot in practise is collaborative filtering. Collaborative filtering is the technique that uses the measured preferences of a group of users to predict the preferences of a new user (Herlocker et al., 2004). A how explanation can infer this through a sentence relating to the definition above:
The following recommendations are made by using the preferences of users with similar search and buying behavior.
4.2.2 Why explanation. The why explanation is used to show that an RA is designed with the purpose of satisfying the needs and preferences of the user (W. Wang & Benbasat, 2007). In practise this comes down to indicating why a recommendation is made through linking it with customer preferences. Since operating systems are among the main criteria of a consumer for choosing a smartphone (Nielsen, 2012), this product characteristic is chosen as customer preference.Therefore the following why explanation is given:
23
4.2.3 Trade-off explanation. Trade-off explanations are used to give the user informational guidance aimed at helping users make trade-offs. A text block in the recommendation will indicate one pro and one con of the combination of attributes presented. For example:
- Pro: This smartphone has a clear 5-inch screen
- Con: With the 5-inch screen it will be hard to reach the corner of the smartphone when
using it with one hand.
The trade-offs are constant between both the group of unfamiliar and familiar phones to minimize measurement bias.
4.2.4 Familiar products. Familiarity is an individual’s understanding of an entity, in many cases based on previous experience. In this research the familiarity will be inferred from a recommendation, thus from a product presented. Because it is not possible to know whether an individual perceives a product as familiar up front, without using a longitudinal research, for this variable products will be used that are in general seen as familiar and compared to products that are generally seen as unfamiliar. Smartphones will be used as product and are selected through a pre test later discussed in this chapter. Familiarity can occur both with negative and positive associations, and for familiarity to have an effect on initial trust it has to be with a positive association, as defined in H5. Therefore the respondents are asked two questions namely ‘the smartphone is familiar to me’ and ‘the smartphone gives me a feeling of goodwill’ both questioned on a seven point Likert scale (disagree-agree) (Ha & Perks, 2005). Every choice will consist out of three products which will vary in familiarity. Therefore there are four different levels of familiarity, namely from zero to three levels. Smartphones are used as product in this research since (1) smartphones are widely adopted1 and (2) smartphones are offered in combination with a subscription, which lasts to up to two years. When using subscriptions a consumer will therefore go through the purchasing process frequently (every one or two years), increasing the probability of having more general knowledge on smartphones and therefore of being a product expert.
1
24
4.2.5 Detailed product information. Detailed information will be presented in both pictures and text. The two levels are placed at the same place and have approximately the same size, since the effect of this variable is based on heuristics and therefore bias is avoided as much as possible. To create a believable situation, the text and pictures are different for every phone presented.
4.2.6 Links to independent reviews. Expert reviews and other customers’ evaluations are both a
display of web links. Because in this research only the display in the output of the RA is measured, these links do not actually work, but still it is expected that through presence it will infer trust
4.3 Choice Based Conjoint Design
The design of the CBC is based on a design randomly generated by Sawtooth SSI web 6.6.6. (Appendix II). Two questionnaires are created of each 8 choice sets and a hold out choice set. Also within each choice set three profiles has to be evaluated. These numbers are set by conducting a advanced design efficiency test in Sawtooth SSI Web 6.6.6 (Appendix II). This test estimates the parameters under aggregate estimation, based on both the sample size (here assumed 200) and elements of design efficiency. Firstly the test simulates random respondent answers to the questionnaire, after which a multinomial logit is performed. The standard error found in the outcome reflects the precision obtained for each parameter. The number of questions was defined by taking the minimum of choice sets that created an standard error of below 0.05, as to minimize fatigue with the respondent and still keep the precision for each parameter high enough. Previous literature also showed that respondents can reliably evaluate up to 20 choice set (Johnson & Orme, 1996) meaning this research (9 choice sets) has a feasible amount of choice sets.
25
Below the hold out question is presented, the difference with the other questions is that this question is fixed for every respondent.
Figure 2: hold out task
26 4.4 Product expertise
Product expertise is questioned by using a self-report measure including four questions (Alba & Hutchinson, 1987). The respondents are classified on their own perceptions of expertise rather than objective measures, because the idea of their own competence may be more relevant to their trust in RA recommendations (Wagner et al., 2003). Also these two ways of measuring product expertise has been shown to converge (Mitchell & Dacin, 1996). On a 7-point likert scale the respondents were asked ‘how familiar are you with smartphones? (not familiar – extremely familiar), ‘how clear an idea do you have about which characteristics of a smartphone are important in providing you maximum usage satisfaction?’ (not very clear – very clear), ‘I know a lot about smartphones’ (agree-disagree) and ‘how would you rate your knowledge about smartphones relative to the rest of the population?’(one of the most knowledgeable – one of the least knowledgeable).
4.5 Comfort with computers and online shopping
The questions ‘how comfortable are you with computers?’ and ‘How comfortable are you with online shopping?’ (seven point likert scale) are included in the survey. These are descriptive statistics used in a paper investigated the trust in recommender systems (Komiak & Benbasat, 2006). Since a level of comfort is needed for trust to occur, this is a relevant question to include in the survey. Also as product or service usage can be related to heterogeneity in consumer preferences (Green & DeSarbo, 1979), these variables could be an effective in segmenting in a later stage.
4.6 Demographics and psychographics
27 4.7 Familiarity Pre Test
To determine the correct smartphones to use for familiarity a pre test is conducted. Ten smartphones are chosen on similarity, six of them sold in Europe and therefore expected to be familiar and four of them only sold outside Europe and therefore expected to be unfamiliar. Because the smartphones will have to relate to a positive familiarity there are both questions regarding familiarity and goodwill. 22 respondents are asked to fill out the survey. The results are displayed in table 2 below. The highest familiarity and goodwill are found for Samsung Galaxy S4, Iphone 5 and HTC One. Conversantly the lowest familiarity and goodwill are found for Neken N6, Jiayu G5 and Mysaga. An LSD test in Appendix III shows that the individual smartphones of both groups highly significantly differ from each other (p<0.001), that is between groups, in both familiarity and goodwill.
Familiarity (Mean) Goodwill (Mean) Samsung Galaxy S4 4.18 4.00 LG G2 2.00 2.23 Neken N6 1.41 1.86 Lenovo S960 VIBE X 1.55 1.91 Iphone 5 4.18 4.05 HTC one 3.77 3.36 Jiayu G5 1.41 1.73 Sony Xperia Z 3.55 3.23 Nokia Lumia 1020 2.59 2.27 Mysaga M2 1.36 1.59 n=22 Bold=chosen smartphones
Table 2: Means pre-test
4.8 Data collection
28
respondents. As seen in table 3 the majority has experience with an RA and over 70% purchases more than 5 times a year.
Experience Mean Distribution
Frequency online shopping per year 3.23 2 = 1-4 times 28.2% 3= 5-8 times 32.7% 4= 9-12 times 26.7% 5 = >12 times 12.4% Frequency usage RA 2.58 1 = Never 29.2% 2 = 1-2 times 38.1% 3 = 3-4 times 10.9% 4 = >4 times 21.8% n=202
Table 3: experience with online shopping and RA's
4.9 Data Analysis
4.9.1 Factor Analysis. The primary purpose of a factor analysis is to check whether there is an underlying structure among variables (Hair et al., 2010) Since product expertise is surveyed using four questions a factor analysis is performed to define whether there is an underlying structure. Both the KMO test (>0.8) and the Bartlett’s Test of Sphericity (p<0.05) indicate that factor analysis is appropriate. The Chronbach’s Alpha is very high (0.923), and deleting any variable would decrease the Chronbach’s Alpha (Appendix IV). In order to define the number of factors, the Principle Component Analysis is used, which uses the total variance and derives factors that contain proportions of unique variance (Hair et. al., 2010). One component is found as seen in Appendix IV, making one factor for product expertise the best option. Product expertise will be used as an active covariate, to be able to distinguish high from low product expertise in the latent class analysis.
29
Averages manipulation check
Familiarity Goodwill Familiar smartphones Samsung Galaxy S4 5.47 4.85
HTC One 4.89 4.15 Iphone 5 5.45 4.88 Unfamiliar smartphones Jiayu 1.34 2.43 Neken N6 1.34 2.41 Mysage M2 1.37 2.45 n=202
Table 4: Averages manipulation check
4.9.3 Choice Based Conjoint Analysis. A conjoint analysis is developed to get insight in how respondents develop preferences for an object. It assumes that respondents value an object by combining the value of each separate attribute (Hair et al. 2010). This study uses CBC, which presents profiles in sets and enables to estimate at aggregate, segment and individual level (Hair et al. 2010). In this research aggregate and segment level are investigated. This will be done by using the software Latent Gold 4.5, which can identify latent classes (segments).
Estimating choice based conjoint models, or discrete choice models, is done through the maximization of utility with the use of multinomial logit (Elrod, Louviere, & Davey, 1992). Including the variables previously described the following model is defined, for each segment i.
= + + + +
In utility function above, U represents the utility, , , and represent the part-worth values, which are related to the level j(j=1,2,3); k(k=1,2,3); l(l=1,2,3,4); m(m=1,2,3,4), for the attributes indrev = independent reviews, detinf = detailed information, fam = familiarity and expl = explanations.
30
In order to check whether a familiarity as a linear function improves the model and to chose the appropriate number of latent classes, the model fit is compared. For the model fit both the Bayesian Information Criterion (BIC) and the Consistent Aikaike Information Criterion (CAIC) are used, since they allow for comparison of non-nested models. These criteria seek to use the considerations of both the ‘best’ approximation to reality and the accuracy of the estimation by using the natural logarithm of the likelihood L to estimate model fit. These criteria also penalize for a lack of parsimony by taking in account the number of parameters and observations, but CAIC penalizes more than BIC (Bozdogan, 1987).
31 0 20 40 60 80 100 1 2 3 4 5 6 7 Comfort with computers Comfort online shopping 5. RESULTS 5.1 Descriptive statistics
The descriptive statistics of the residual respondents are displayed table 5 below. There is a nice distribution for gender and education. Some bias is found in region, age, occupation and income, meaning that the distribution is focused on a few groups or one group. In region most of the respondents answered middle (18.8%) north (39.6%) and east (31.7%). Age shows a bias towards people between 20-24 (51.5%), while occupation show the highest percentage in students (33.2%) and Full-time job (32.7%). Income above €2501 is represented by 5.4% while the other categories are nicely distributed.
Descriptive statistic % Descriptive statistic %
Gender Female 49.5% Education MAVO/VMBO 1.5% Male 50.5% HAVO/VWO 7.4% MBO 18.8% Region Netherlands Middle 18.8% HBO 39.6%
North 39.6% WO 32.7%
East 31.7%
South 3.5% Occupation Student 33.2% West 3.5% Student with a job 15.3%
None 3% Unemployed 4%
Part-time job 14.4% Age <20 8.9% Full-time job 32.7% 20-24 51.5% Retired 0.5% 25-29 16.8% >29 22.8% Income <€500 29.2% €500-€1500 36.1% €1501-€2500 21.3% €2501-€3500 5.4% Not answered 7.9%
Table 5: Descriptive statistics
The distribution among the respondents of both comfort with computers as comfort with online shopping as seen in figure 3 is skewed towards high comfort, where he majority answered 5, 6 or 7.
32 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 U ti li ty
Familiarity
5.2 Choice-Based Conjoint analysis
5.2.1 Aggregate results. In table 6 the results of the aggregate model are displayed. Based on the Wald statistic all the attributes are highly significant (p<0.001). The relative importance indicate the importance of the individual attribute, compared to the other. Detailed information is found most important by the respondents in the survey (31.09%), followed by explanation and link, which are of about equal importance (28.07% and 26.05%). Familiarity is found least important by the respondents with a relative importance of 14.79%. According to the utilities, the recommendation agent with the combination of a picture, a consumer link, 3 familiar products and a trade-off explanation will be perceived as most trustworthy.
Attributes Levels Wald β Relative importance
Explanations* No explanation 283,57 0 28.59% How explanation 0.1922
Why explanation 0.1558 Trade-off explanation 1.3452
Familiar products* No familiarity 56.48 0 13.93% 1 familiar product 0.1793
2 familiar products 0.5399 3 familiar products 0.6556
Detailed information* No detailed information 299.73 0 31.64%
Text 0.9174 Picture 1.4886 Links to independent reviews* No link 221.93 0 25.84% Expert link 0.7486 Consumer link 1.2157 * p<0.001
Table 6: Parameters aggregate model
5.2.2 Model fit aggregate model. Due to the ordered nature of the levels of familiarity, the attribute is tested for linearity. Figure 4 suggests a linear relationship through the linear form. Therefore two models are estimated, one including familiarity as a linear model and including familiarity as a nominal variable. These are both compared by model
33
Model BIC(LL) CAIC(LL) df
Aggregated model with nominal familiarity 2616.337 2626.337 192 Aggregated model with numeric/linear familiarity 2608.464 2616.464 194
Table 7: linear versus nominal familiarity compared
Table 7 indicate a lower BIC and CAIC for the aggregated model with familiarity as a linear function. Therefore the model fit is highest when the attribute familiarity is entered linear.
5.2.3 Hypothesis testing on aggregate results. Table 8 shows the results after estimating the model with a linear familiarity parameter, including the z values for the individual levels of the attributes. The z value indicates whether the effect significantly differs from the reference category, here seen as the no-levels. All the levels but 'why explanation' show a significant difference effect (│z-value│<1.96 = p<0.05). Also the utilities of these levels are positive, meaning they represent a significant increase over the no-options. These results therefore indicate that H1 (how explanations), H3 (trade-off explanations), H5 (positive familiar products), H6 (detailed information) and H7 (independent reviews) are supported. Why explanation has a z value of 1.3784 with a corresponding p value of 0.16758, indicating no significance, thus H2 is not supported. Familiarity has no reference category since it is linear, therefore the z value indicates that the effect is significantly larger than a zero effect.
Attributes Levels Z value Wald β Relative importance
Explanations No explanation - 287.65 0 28.07% How explanation* 1.9841 0.1928
Why explanation 1.3784 0.1301 Trade-off explanation 14.3238 1.3255
Familiar products 7.3543 54.08 0.2328 14.79% Detailed information No detailed information - 300.23 0 31.09%
Text* 10.0667 0.9033 Picture* 17.1687 1.4681 Link to independent reviews No link - 231.36 0 26.05% Expert link* 9.4858 0.7573 Consumer link* 15.1892 1.2298 * p<0,05
34 0 0,2 0,4 0,6 0,8 1 1,2 1,4 U ti li ty
Explanations
Also the tradeoff explanation has the largest utility within explanations, as seen in figure 5. To check whether H4 (how explanation/why explanation<trade-off explanation) is supported a model is estimated with the trade off explanation level as reference category. This allows to compare the other variables relative to the trade off explanation. Table 9 shows that all possible explanations have a negative z value <-1.96 meaning that trade-off explanation has a significantly higher effect on utility compared to all other explanations (p<0.05). Therefore H4 is supported.
Explanations with trade-off set to 0
Utility Z value
no explanation -1.3255 -14.3238 how explanations -1.1327 -12.1718 why explanation -1.1955 -13.2693
trade off explanation 0 . Bold = │z-value│<-1.96 = p<0.05
Table 9: Explanations with trade-off as reference category
5.2.4 Selection of number of segments. In this section a latent class analysis is performed. First a check is done whether possible active covariates can improve the model fit, by using an explorative approach. Stepwise the variables gender, age, region, income, occupation, education, frequency of RA usage, frequency of online purchasing, comfort with computers and comfort with online shopping are included and excluded as active covariates in a model. In this model product expertise is already included as active covariate, since investigating that variable is one of the main aims of this study.
The significance level of the possible active covariates are checked. The, already included, active variable product expertise is mostly significant (p<0,05). For the exploratory variables only frequency of RA usage appears to be significant (p<0.05) in some, but not all of the models. To
35
check whether the covariate frequency of RA usage, increases the model fit, both the BIC and CAIC are compared. As seen in table 10 from the 2-Class model on, the BIC and CAIC are structurally lower in the product expertise only models. Therefore it is chosen to only include product expertise as a covariate for classifying the segments.
Product expertise only Product expertise and frequency usage of RA BIC(LL) CAIC(LL) df Class.Err. BIC(LL) CAIC(LL) df Class.Err.
1-Class model 2608.464 2616.464 194 0 2608.464 2616.464 194 0 2-Class model 2482.276 2500.276 184 0.0588 2487.478 2506.478 183 0.059 3-Class model 2435.987 2463.987 174 0.0687 2445.299 2475.299 172 0.0699 4-Class model 2415.867 2453.867 164 0.1001 2422.278 2463.278 161 0.0926 5-Class model 2412.278 2460.278 154 0.1146 2421.221 2473.221 150 0.1081 6-Class model 2407.743 2465.743 144 0.1004 2424.11 2487.11 139 0.1013 7-Class model 2425.901 2493.901 134 0.0923 2443.434 2517.434 128 0.0948
Table 10: Product expertise only compared to product expertise and frequency usage of RA
Second the number of segments are selected, again by investigating the BIC and CAIC. Table 11 below, presents the lowest BIC for the 6 class model, while the lowest CAIC is presented for the 4-class model. The distribution of the size is similar, with an emphasis on one large segment. The Classification error shows in which model the classification is best, which indicates how well each respondent is classified into classes, given the observed data. This is similar for both models (0.1001 versus 0.1004).
Model BIC(LL) CAIC(LL ) Df Class. Err. Size Class 1 Size Class 2 Size Class 3 Size Class 4 Size Class 5 Size Class 6 Size Class 7 1-Class model 2608.464 2616.464 194 0 100% 2-Class model 2482.276 2500.276 184 0.0588 66.36% 33.64% 3- Class Model 2435.987 2463.987 174 0.0687 57.31% 33.59% 9.10% 4- Class Model 2415.867 2453.867 164 0.1001 49.76% 23.67% 17.18% 9.39% 5-Class Model 2412.278 2460.278 154 0.1146 41.18% 19.17% 15.59% 15.16% 8.9% 6-Class Model 2407.743 2465.743 144 0.1004 39.94% 17.52% 15.11% 9.94% 8.91% 8.58% 7-Class Model 2425.901 2493.901 134 0.0923 41.25% 12.18% 10.93% 10.88% 10.61% 10% 4.16% Bold = lowest BIC or CAIC score
Table 11: Model fit multiple classes
36 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8
Class 1 Class 2 Class 3 Class 4
Informationprov Links
Familiarity Explanations
other hand are rather similar, making it hard to distinguish the segments. Therefore the 4-class model is chosen.
4-Class model 6-Class model
Class1 Class2 Class3 Class4 Class1 Class2 Class3 Class4 Class5 Class6
Explanations 0.3523 0.1186 0.2944 0.0915 0.3042 0.4786 0.1242 0.1787 0.1022 0.2801 Familiar products 0.119 0.095 0.3391 0.1186 0.1025 0.0165 0.3094 0.5039 0.1241 0.0966 Detailed information 0.2724 0.5543 0.1402 0.1006 0.3011 0.2941 0.5066 0.1787 0.0734 0.2314 Links to independent reviews 0.2563 0.2321 0.2263 0.6894 0.2921 0.2107 0.0598 0.1387 0.7003 0.3919 Bold = highest score for attribute
Tabel 12: relative importance of 4-class model compared to 6-class model
5.2.5 Results segment analysis. The relative importance of the four segments are displayed in figure 6 below. The Wald (=) statistics found significant for detailed information, links and explanations and familiarity (p<0.05), meaning that the parameters significantly differ over segments. Class 1 appear to prefer explanations (35,23%), Class 2 detailed information (55,43%), class 3 familiarity (33.91%) and class 4 links (68,94%), since the relative importance is highest for these attributes within these classes.
37
Class 1 Class 2 Class 3 Class 4
R²(%) 49.19% 58.91% 21.61% 73.58% N(%) 49.76% 23.67% 17.18% 9.39% Relative importance Explanations 35.23% 11.86% 29.44% 9.15% Familiar products 11.9% 9.5% 33.91% 11.86% Detailed information 27.24% 55.43% 14.02% 10.06% Links to independent reviews 25.63% 23.21% 22.63% 68.94%
Parameter values β z-value β z-value β z-value β z-value
no explanation* 0 . 0 . 0 . 0 . how explanations 0.3244 1.5647 0.3536 0.9264 0.3495 1.071 0.301 0.495 why explanation -0.1795 -0.9419 0.1065 0.35 0.4516 1.4358 -0.3695 -0.654 trade off explanation 2.3486 12.0164 0.7451 1.8331 1.2186 3.8431 -0.0978 -0.1723 Familiar products 0.2847 3.9684 0.199 1.5807 0.468 5.1122 -0.2896 -1.1787 no detailed information* 0 . 0 . 0 . 0 . text 1.9547 9.7217 1.26 2.9964 0.0889 0.3681 0.3146 0.6098 picture 1.8617 8.0841 3.4839 8.2093 0.5805 2.2234 0.7367 1.0936 no links* 0 . 0 . 0 . 0 . Expert link 0.9324 5.1295 1.4589 3.8981 0.9368 4.6138 2.5601 2.3106 Consumer link 1.8396 8.0227 1.4307 3.8539 0.1188 0.4837 5.0505 4.5104 Intercept 0 . -0.9385 -2.6442 -1.0683 -3.5677 -1.7236 -5.1709 Product Expertise 0 . 0.8833 2.6895 0.2936 1.1161 -0.2327 -0.8935
a. * = The parameters are set to zero because they are redundant (dummy first code) b. Intercept and Product Expertise are both significant (p<0.05)
c. Bold = │z-value│>1.96 = p<0.05
Table 13: Parameters 4-class model
Table 13 presents the parameters and corresponding z values for the four segments. The parameter (β) is a measure of the influence on the utility of the attribute, compared to the reference category, which in all cases are the ‘no options’. A positive β therefore can be interpreted as an increase over doing nothing, and vice versa for a negative β. As seen by the z values, the significance varies across segments (p<0.05). A further profiling of classes is given below:
Class 1 – Trust sensitive users: This segment represents 49.46% of the respondents. Because
38
explanations has the largest effect on initial trust (β = 2.34) followed by text (β = 1.95), picture (β = 1.86), consumer link (β = 1.83), expert link (β = 0.93) and familiarity (β = 0.28).
Class 2 - Product experts: This segment represents 23.67% of the respondents. The segment
has a significantly higher probability of having product expertise compared to the other segments. Within this segment there is clearly a strong preference for detailed information (55.43%). For this segment a picture has the largest significant effect on initial trust (β = 3.48) followed by expert link (β= 1.45), consumer link (β= 1.43) and text (β = 1.26).
Class 3 - Familiarity seekers: This segment represents 17.18% of the respondents and has the
highest preference for familiarity (33.91%). Therefore they are referred to as familiarity seekers. For this segment familiarity has the largest significant effect on initial trust (β = 1.21) followed by expert link (β = 0.93), picture (β = 0.58) and familiarity (β = 0.46).
Class 4 – Opinion seekers: This segment represents 9.39% of the respondents, making it the
smallest segment. There is a very strong preference for independent links (68.4%). Since the independent links represent opinions of others the segment can be classified as opinion seekers. For this segment expert link has the strongest significant effect on initial trust (β = 5.05) followed by consumer link (β = 2.56).
39
55.2.6 Hypothesis testing on segment level. Based upon the z values in table 13 (p<0.05) the conclusions below in table 15 are made. Because class 2 has a higher probability of having product expertise this class is indicated as ‘high product expertise’ while class 1,3 and 4 are indicated as ‘low product expertise’. By looking at the patterns across the classes the conclusions for H8, H9 and H10 can be made.
Table 5: Hypothesis testing on segment level
For class three must be noted that detailed information and links to independent reviews are not supported, since for both attributes only one of the two levels indicate a significant improvement over the non option. In order for a hypothesis to be supported, both the levels have to indicate a positive significant effect.
H1 is not supported in all of the classes, while it is supported on aggregate level. H2 is not supported over all classes, therefore also H8 is not supported. H3 is supported for class 1 and 3, since these classes show both a positive significant effect (p<0.05) on initial trust. Therefore, due to the fact that both how and why are not supported over all classes, H4 is also supported for class 1 and 3. Class 1 and 3 also shows a positive significant effect (p<0.05) for familiar products, supporting H5.
Even though text is most valued by a low product expertise group, H9 is not supported. Reason for that is that the other level of detailed information, picture, is valued most by the high
Lowproduct expertise High product expertise Low product expertise Low product expertise Aggregate model
Class 1 Class 2 Class 3 Class 4
H1: How explanation Supported Not supported Not supported Not supported Not supported H2: why explanation Not supported Not supported Not supported Not supported Not supported H3: Trade off explanation Supported Supported Not supported Supported No supported H4: How explanation/Why
explanation < Trade-off explanation
Supported Supported Not supported Supported Not supported H5: Familiar products Supported Supported Not supported Supported Not supported H5: Detailed information Supported Supported Supported Not supported Not supported H6: Links to independent reviews Supported Supported Supported Not supported Suported
H8: Why explanation Not supported
H9: Detailed information Not supported
40
expertise segment. H10 is not supported even though for class four both expert link and consumer link have the strongest positive significant effect on initial trust (p<0.05). This is due to the fact that the expert links in class 1 and three (low product experts) have a positive significant effect (p<0.05), but to a lower extent than the high product expert segment. This inconsistent pattern thereby rejects H10.
5.2.7 Predictive validity. In order to validate the predictive power of the estimated models, both
the aggregated and the 4-Class model are compared with the hold out sample. Table 14 below shows that the 4-class model performs better for the in-sample (79% versus 65.79%) as well as for the out-sample (54.59% versus 41.58%). Both the in- and out-sample perform better than a random prediction (33.33%), thus the samples validate that the 4-class model has better predictive power than a random choice prediction.
Bold = Hit Rate
Table 14: Hit rate in-sample versus out-sample
Aggregated model 4-Class model
Hit Rate Hit rate
41
6. DISCUSSION
6.1 Conclusion
This study attempts to create an insight in how the initial trust of a user in RA’s can be increased by investigating to what extent do explanations, familiar products and information about the product, in the output of an RA, have an impact on a users’ initial trust in the RA. Also to create a more specific context, consumer product expertise is taken into account by investigating whether the effect of RA output on initial trust differs for product experts versus product novices. In this section these concepts are discussed.
In this study three explanations are investigated, namely how, why and trade-off explanations. Estimating the utility through an aggregate model concludes that both how and trade-off explanations have a positive effect on initial trust, while including a why explanation did not have a significant effect. Because a how explanation creates insight into the process of an RA, a possible explanation is that a consumer will trust the RA’s abilities more. This relates to competence belief of a consumer (McKnight & Chervany, 2002) A trade-off explanation on the other hand will give the user informational guidance to help users make trade-offs between products and attributes, and therefore the consumer possibly perceives the RA to be impartial, relating to the integrity belief (Mayer et al., 1995). Trade-off explanations seems to have significantly and to a big extent the largest effect on initial trust, which is consistent with the classification based on Toulmin’s model of argumentation (Gregor & Benbasat, 1999). It is namely proposed that trade-off explanations conform more to Toulmin’s model compared to the how and why explanations and therefore is more convincing and creates more initial trust.
42
Two kinds of detailed information are investigated, namely the inclusion of a picture and the inclusion of a text describing the product. When estimating the aggregate model these both have a significant positive effect on initial trust. This is consistent with the notion that providing a useful interface will create the a feeling that the RA cares about the user, creating a benevolence belief (McKnight & Chervany, 2002). The inclusion of a picture in an RA has a larger effect than the inclusion of a text describing the product on initial trust. A possible explanation is given by processing fluency literature that indicate that stimuli that can be more easily processed generally inspire more favourable attitudes (Reber, Schwarz & Winkielman, 2004).
Two different links to independent reviews are defined, namely a link to expert reviews and a link to consumer reviews. These both appear to have a positive significant effect on initial trust when included in an RA. This can relate to the sense that the RA both cares about the user and is unbiased, thus affect the benevolence and integrity belief (McKnight & Chervany, 2002). The web link to consumer reviews has a stronger effect than the web link to expert reviews. This could possibly be accounted to the fact people tend to form stronger trusting beliefs towards other people that are more similar to them, because they tend have the same goals and values (McKnight et al., 1998) As other consumers are more likely to have the same (consumer) goals as the respondent than the experts, a link to consumer reviews shows a stronger effect.
43
As previous research failed to focus on the output of an RA, this study contributes to the existing literature by focusing on the effects of the output of an RA on initial trust. Previous research has investigated possible factors affecting a users trust in RAs, but did so by using for example the input section (during elicitation of preferences) of an RA (W. Wang & Benbasat 2007). This is a form of explicitly extracting information from a customer. Such research disregards the fact that RAs can also implicitly gather information from customers, by for example monitoring browsing and buying behaviour (Xiao & Benbasat, 2007). When the customer information is solely implicitly gathered, the output of an RA is the only notable touch point for the consumer with the RA by which the consumer can create evaluations. Therefore the focus on the effects of output of an RA on initial trust is an important and relevant contribution to research.
6.2 Managerial Implications
Consumers are increasingly shopping online, therefore it is becoming very important for vendors to get an understanding in the online behaviour of consumers, especially what triggers consumers into buying. Industry experts indicate that personalizing shopping experience is an effective way for online retailers to turn browser into buyers (Verton, 2001). Recommendation agents are used as a tool to personalize the shopping experience (Häubl & Trifts, 2000), but in order for the consumer to accept the recommendations made by the RA, a certain level of trust has to occur (Häubl & Murray, 2001). Therefore this research is relevant for managers in that it creates insights in how to create RAs that infer initial trust with the consumer.
This research gives insight in what elements can be used by the manager to create initial trust with the consumer. The three, to an equal extent, most valued attributes are detailed information, links to independent reviews and explanations. Including all of these elements may make the output to large, therefore the manager can make a decision based on the optimal combination of levels of these attributes. The optimal combination would be to include a picture, a link to other consumers’ reviews and a trade-off explanation, since these are the levels with the highest effect on initial trust.
