Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Investigating consumers' use of product recommendation agents: understanding the influence of product… Lee, Young Eun 2007

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2007-318119.pdf [ 9.4MB ]
Metadata
JSON: 831-1.0100651.json
JSON-LD: 831-1.0100651-ld.json
RDF/XML (Pretty): 831-1.0100651-rdf.xml
RDF/JSON: 831-1.0100651-rdf.json
Turtle: 831-1.0100651-turtle.txt
N-Triples: 831-1.0100651-rdf-ntriples.txt
Original Record: 831-1.0100651-source.json
Full Text
831-1.0100651-fulltext.txt
Citation
831-1.0100651.ris

Full Text

I N V E S T I G A T I N G C O N S U M E R S ' U S E O F P R O D U C T R E C O M M E N D A T I O N A G E N T S : U N D E R S T A N D I N G T H E I N F L U E N C E O F P R O D U C T T Y P E A N D I N - S T O R E C O N T E X T by Young Eun Lee M.B.A. , Yonsei University, 2001 B.A. , Yonsei University, 1997 A THESIS SUBMITTED IN PARTIAL F U L F I L L M E N T OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in The Faculty of Graduate Studies (Business Administration) THE UNIVERSITY OF BRITISH C O L U M B I A September 2007 © Young Eun Lee, 2007 ABSTRACT Product Recommendation Agents (RAs) are web-based software systems that advise consumers about what to buy based on the needs expressed by those consumers. Most research on RAs has focused on evaluating the different algorithms that generate recommendations, while the effectiveness of RAs is determined by many factors beyond design of algorithms. This dissertation focuses on two important but understudied factors, namely, the influences of product type and in-store context. These factors were investigated in two separate laboratory experiments: Study 1 and Study 2. Study 1 examined the influence products with high emotional contents on consumers' use of RAs. Such products are characterized by attribute conflicts which involve the correlation of the favourable values of some attributes with the unfavourable values of others. Two RAs - one that highlights attribute conflicts and one that presents them implicitly - were developed and compared. The experiment results show that the R A highlighting attribute conflicts negatively influences consumers' perceptions and acceptance of RAs, as compared to the R A obscuring the conflicts. In addition, task emotionality - the degree to which a decision task is perceived to bring severe negative consequences - moderates such relationships. Study 2 examined the influence of in-store contexts on consumers' use of RAs. The in-store context has become important with the advent of mobile RAs operated on handheld devices, such as Personal Digital Assistants. In the in-store context, compatibility between the way the store displays the products and the way RAs guides consumers' decision making is an important predictor of decision performance. Two RAs which guide consumers' decisions in two different manners - alternative-driven and attribute-driven ways - were contrasted. Due to a limit to displaying complex products sorted by every attribute on product shelves, products are displayed randomly by ii alternative in a store. Therefore, the alternative-driven R A is more compatible with the store's product displays than attribute-driven RA. The experiment results show that compatibility increases both accuracy of decisions and perceived control. In addition, mobile R A use, as compared to R A non-use, decreases consumers' perceived effort and increases their intentions to return to the store. iii TABLE OF CONTENTS ABSTRACT ii TABLE OF CONTENTS iv LIST OF TABLES vii LIST OF FIGURES ix CHAPTER 1: INTRODUCTION 1 1.1 RESEARCH OBJECTIVES AND QUESTIONS 6 1.2 METHODOLOGY AND MAJOR FINDINGS .7 1.3 STRUCTURE OF T H E DISSERTATION 8 CHAPTER 2: LITERATURE REVIEW 10 2.1 DECISION STRATEGIES 10 2.1.1 Properties of Decision Strategies , 13 2.1.2 Comparisons between WAD and EBA , 15 2.1.3 Comparisons between ADDIF, EQW, M C D , SAT and WAD, EBA 15 2.2 CONSTRUCTIVE DECISION MAKING THEORY 16 2.3 EFFECTS OF ATTRIBUTE CONFLICTS ON CONSUMERS' PERCEPTIONS AND ACCEPTANCE OF RECOMMENDATION AGENTS 16 2.4 MOBILE RECOMMENDATION AGENTS FOR IN-STORE DECISION MAKING 22 2.5 TABLES , 25 CHAPTER 3: EFFECTS OF ATTRIBUTE CONFLICTS ON CONSUMERS' PERCEPTIONS AND ACCEPTANCE OF RECOMMENDATION AGENTS 27 3.1. INTRODUCTION 27 3.2 THEORY DEVELOPMENT 31 3.2.1 RA-WAD and RA-EBA 31 3.2.2 Perceived Control 33 3.2.3 Perceived Quality of Recommendations 35 3.2.4 Perceived Effort 36 3.2.5 Usage Intentions 37 3.2.6 Moderator Variable: Task Emotionality 39 3.3 R E S E A R C H M E T H O D 42 3.3.1 RA Design .' 42 3.3.2 Alternatives and Attributes 43 3.3.3 Manipulation of Task Emotionality 44 3.3.4 Operationalization of the Dependent Variables 47 3.3.5 Participants and Experimental Procedures 47 3.4 R E S U L T S A N D A N A L Y S E S 49 3.4.1 Manipulation Checks 49 3.4.2 Testing of Hypotheses 50 3.4.3 Relationships among the Dependent Variables 53 3.4.4 Structural Equation Model Analysis 55 3.5 DISCUSSION A N D C O N C L U S I O N S 58 3.5.1 Summary of Findings 58 3.5.2 Discussion of the Results 58 3.5.3 Limitations 60 3.5.4 Contributions and Implications 61 3.5.5 Future Research 62 3.6 T A B L E S A N D F I G U R E S 64 CHAPTER 4: EFFECTS OF COMPATIBILITY ON CONSUMERS' IN-STORE DECISION MAKING WITH MOBILE RECOMMENDATION AGENTS 76 4.1 I N T R O D U C T I O N 76 4.2 T H E O R Y D E V E L O P M E N T 80 4.2.1 Use of Mobile RAs for In-Store Decision Making 80 4.2.2 Alternative-Driven R A vs. Attribute-Driven R A 84 4.2.3 Effects of Guidance Directions on Decision Processes • 87 4.2.4 Compatibility between the R A ' s Guidance Directions and the Store's Product Displays 88 4.2.5 Effects of Compatibility on Decision Accuracy, Perceived Effort, Perceived Control, and Usage Intentions 90 4.3 R E S E A R C H M E T H O D 92 4.3.1 Alternatives and Attributes 93 4.3.2 R A Design : 94 4.3.3 Computing Devices 96 4.3.4 Operationalization of the Dependent Variables 96 .4.3.5 Experimental Tasks 98 4.3.6 Desirable vs. Less Desirable Printers 99 4.3.7 Store Design 101 4.3.8 Participants and Experimental Procedures 103 4.4 R E S U L T S A N D A N A L Y S E S 105 4.4.1 Overview 105 4.4.2 Control vs. R A groups 106 4.4.3 R A - A L vs. R A - A T Groups 108 4.5 DISCUSSION A N D C O N C L U S I O N S 110 4.5.1 Summary of Findings 110 4.5.2 Discussion of the Results 111 4.5.3 Limitations 112 4.5.4 Contributions and Implications 115 4.5.5 Future Research 117 4.6 T A B L E S AND FIGURES 119 CHAPTER 5: CONCLUSIONS AND FUTURE RESEARCH 136 5.1 CONCLUSIONS 136 5.2 LIMITATIONS 138 5.3 CONTRIBUTIONS 140 5.4 F U T U R E R E S E A R C H 142 REFERENCES 143 APPENDICES 151 A.1 N A T U R A L GOMS L A N G U A G E ANALYSES 151 A. 1.1 Rationale for Conducting NGOMSL Analyses 151 A. 1.2 Predicting Execution Time with NGOMSL Analyses 152 A. 1.3 NGOMSL Analyses 153 A. 1.4 Summary .• .' 154 A.2 HIGH EMOTION DECISION T A S K 156 A.3 L O W EMOTION DECISION T A S K 157 A.4 E X P E R I M E N T A L TASK FOR T H E RA GROUPS 158 A.5 E X P E R I M E N T A L TASKS FOR T H E C O N T R O L GROUP 160 A.6 T A B L E S 162 vi LIST OF TABLES Table 2.1 Properties of the Six Decision Strategies (Payne et al. 1993, page 32) 25 Table 2.2 Summary of Previous Research on Attribute Conflicts 26 Table 3.1 Attributes of Used Cars 64 Table 3.2 Dependent Variables 65 Table 3.3 Demographics of Participants.. 66 Table 3.4 Means, SDs, Inter-Construct Correlations and Average Variance Extracted... 66 Table 3.5 Descriptive Statistics 66 Table 3.6 A N C O V A Results: Perceived Control 67 Table 3.7 A N C O V A Results: Recommendation Quality 67 Table 3.8 A N C O V A Results: Perceived Effort 68 Table 3.9 A N C O V A Results: Usage Intentions 68 Table 3.10 Factor Loadings and Cross Loadings 69 Table 4.1 Attributes of Printers 119 Table 4.2 Dependent Variables 120 Table 4.3 Decision Tasks and Desirable Printers 121 Table 4.4 Attribute Levels of 22 Printers 122 Table 4.5 Overview of Hypotheses Testing 123 Table 4.6 Descriptive Statistics 123 Table 4.7 Printers Chosen for a Family Member 124 Table 4.8 Printers Chosen for a Friend....; 124 Table 4.9 Chi-Square Test Results 125 Table 4.10 A N C O V A Results: Perceived Effort 125 Table 4.11 A N C O V A Results: Intentions to Return to the Store 126 Table 4.12 A N C O V A Results: Number of Times Alternatives Were Viewed 126 Table 4.13 A N C O V A Results: Perceived Effort 127 Table 4.14 A N C O V A Results: Perceived Control 127 Table 4.15 A N C O V A Results: Usage Intentions 128 Table A . l Operators Adopted in the Current Study 162 Table A.2 Assumptions of N G O M S L Analyses 162 Table A.3 N G O M S L Analyses of R A - W A D 163 Table A.4 N G O M S L Analyses of R A - E B A 164 Table A. 5 Summary of N G O M S L Analyses and the Estimated Execution Time 165 LIST OF FIGURES Figure 3.1 Feedback-Control Loop (Morris And Marshall 2004) 70 Figure 3.2 Two Stages of RAs 70 Figure 3.3 R A - W A D 71 Figure 3.4 R A - E B A 72 Figure 3.5 Perceived Control 73 Figure 3.6 Recommendation Quality 73 Figure 3.7 Perceived Effort 74 Figure 3.8 Usage Intentions 74 Figure 3.9 Structural Equation Model Analyses 75 Figure 4.1 Diagram of R A - A L 129 Figure 4.2 Screenshots of R A - A L 130 Figure 4.3 Diagram of R A - A T 131 Figure 4.4 Screenshots of R A - A T 132 Figure 4.5 Experimental Store 133 Figure 4.6 Picture and Price Tag of a Printer (Samsung 2020) 134 Figure 4.7 Experimental Procedures 135 i x CHAPTER 1: INTRODUCTION Product Recommendation Agents (RAs) are web-based software systems that can advise consumers about what to buy based on the needs expressed by those consumers (Wang and Benbasat 2007; Xiao and Benbasat 2007). RAs are specifically designed to assist consumers in purchasing complex goods and services online (Grenci and Todd 2002). RAs have the potential to support consumers in reducing information overload and task complexity, and facilitate accurate product choices (Maes, Guttman and Moukas 1999; O'Keefe and McEachern 1998). Since their introduction in the early 1990s by the University of Minnesota's GroupLens Research Project (www.cs.umn.edu/research/GroupLens/index.htmn, RAs have grown substantially more sophisticated and become established as an essential feature of many web storefronts, such as Amazon.com and BestBuy.com (Leavitt 2006). According to Jupiter Research1, the Internet will influence nearly half of total retail sales in 2010, compared with just 27% in 2005. This projection combines total sales transacted online with sales carried out offline but encouraged by online research. Such rapid growth is in part boosted by recommendation technologies (Leavitt 2006). RAs provide information and suggestions about products which transform web browsers into potential buyers (Leavitt 2006). In addition, the R A has, per se, created a new profitable business model that emerged online (Laudon and Laudon 2006), exemplified by successful comparison-shopping websites enabled by recommendation technologies, such as Shopping.com and Bizrate.com (Xiao and Benbasat 2007). Numerous empirical studies have examined the impact of RAs on consumers' purchase decisions (see Xiao and Benbasat [2007] for a detailed review of prior research). 1 http://www.jupitermedia.corn/coiporate/releases/06.02.06-newjupresearch.html 1 To date most research on RAs has focused on developing and assessing the different underlying algorithms that yield recommendations (Cosley, Lam, Albert, Konstan and Riedl 2003; Swearingen and Sinha 2002). However, consumers determine the effectiveness of RAs based upon many factors beyond design of algorithms, including the characteristics of particular RAs, source credibility, tasks, contexts, and product- and user-related factors (Todd and Benbasat 1992; Xiao and Benbasat 2007). It is likely the neglect of such factors that accounts for the equivocal results reported by some studies on the effectiveness of computer-based decision aids (Todd and Benbasat 1992). Research is needed to investigate these relevant, but understudied, factors in order to broaden our understanding of consumers' perceptions and, as a consequence, their acceptance of RAs. This dissertation focuses on two of these understudied factors, namely, the influences of product type and in-store context. These factors are investigated in two separate empirical studies: Study 1 and Study 2. In the context of B2C eCommerce, the nature of any particular product is an important influence on consumers' processes of searching for and acquiring information and making decisions about purchases (Levin, Levin and Heath 2003; McCabe and Nowlis 2001). Consumers' decision-making processes and outcomes, as influenced by product type, in turn affect consumers' perceptions of RAs (Aloysius, Davis, Wilson, Taylor and Kottemann 2006; Fasolo, McClelland, Lange, Haugtvedt, Machleit and Yalch 2005; Wang and Benbasat 2005; Widing and Talarzyk 1993). Clearly, research into product type is essential to better understand consumers' perceptions and acceptance of RAs. In particular, products with high emotional contents are the focus of attention in Study 1. Such products are characterized by attribute conflicts (Luce, Bettman, Payne and Bettman 2001). Attribute conflicts involve the correlation of the favourable values of some attributes with the unfavourable values of others. Attribute conflicts compel consumers to accept less of one for more of the other; such compromise signals a 2 potential loss of the attribute being traded off. Consumers contemplating products with conflicting attributes are threatened by the fact that the attainment of important goals may be blocked by their own choices (Lazarus 1991; Luce et al. 2001). Consumers value the goals associated with decision making that involves attribute conflicts differently than they value the goals involved in general decision making. For decisions concerning products with a fewer attribute conflicts, consumers value two primary goals: (1) to minimize the effort necessary to make the decision and (2) to maximize the accuracy of the decisions (Bettman 1979; Bettman, Johnson and Payne 1990; Bettman, Luce and Payne 1998). Consumers' emphasis on the effort-accuracy goal changes when attribute conflicts enter the picture (Luce 1998; Luce, Bettman and Payne 1997; Luce, Bettman and Payne 1999; Luce, Bettman and Payne 2000). Consumers' primary motivation then becomes to minimize the confrontation with attribute conflicts, which cause intense decisional stress, even though the attainment of this goal interferes with the achievement of the effort-accuracy goal (Luce et al. 2001). From these variations in consumers' most valued goals, one can infer that consumers' perceptions and acceptance of R A s that offer advice about products with attribute conflicts w i l l also differ. For purchases of products with fewer attribute conflicts, consumers favour and positively appraise those R A s that reduce effort and augment decision accuracy, precisely because these R A s enable them to achieve their effort-accuracy goal (Haubl and Murray 2003; Haubl and Trifts 2000; Wang and Benbasat 2005). Conversely, for purchases of products with strong attribute conflicts, consumers may prefer the R A s that hide the conflicts, as these R A s allow them to attain their most prominent goal: to avoid the stressful knowledge that tradeoffs among conflicting attributes must be made (Luce 1998). With this in mind, I developed two R A s in Study 1: one that highlights attribute conflicts, and one that presents attribute conflicts implicitly. The R A designed to 3 highlight attribute conflicts employed a compensatory strategy, while the RA designed to hide attribute conflicts employed a non-compensatory strategy. Decision strategies are the rules decision makers follow in order to solve multi-alternative/multi-attribute preferential choice problems that involve selecting one out of a number of alternatives described by a common set of attributes (Svenson 1979). Decision strategies are an integral part of RAs (Song, Jones and Gudigantala 2007). A number of RAs are implemented based on one (or more) strategies (Song et al. 2007). Moreover, strategies differ in terms of the extent to which they highlight attribute conflicts: Whereas a compensatory strategy presents attribute conflicts explicitly, a non-compensatory strategy obscures them (Bettman et al. 1998; Payne, Bettman and Johnson 1993). Therefore, a RA employing a compensatory strategy highlights attribute conflicts, thereby decreases users' perceived control over manipulating the RA to resolve the conflicts, as compared to a RA employing a non-compensatory strategy. As such, comparison of the two RAs will demonstrate the effects of attribute conflicts on consumers' perceptions and acceptance of RAs. In addition to considering the influence of attribute conflicts on consumers' perceptions and acceptance of RAs, I examine whether or not task emotionality moderates such relationships. Task emotionality is defined as the degree to which a decision task is perceived to bring severe negative consequences (Luce et al. 1997; Luce et al. 1999). Task emotionality determines the extent to which attribute conflicts influence consumers' decision making without altering other cognitive factors associated with the decision problem (Luce et al. 1997; Luce et al. 1999). Therefore, successful detection of the moderating effects of task emotionality counters the alternative hypothesis that factors other than attribute conflicts cause consumers' perceptions of the two RAs to differ from each other. Study 2 examines the influence of in-store contexts on consumers' use of RAs. In-4 store context has become relevant to investigations into RAs with the advent of mobUe RAs. Mobile RAs are operated on mobile handheld devices, such as Personal Digital Assistants (PDAs) and cellular phones (Van der Heijden 2005; Van der Heijden and Sorensen 2002; Van der Heijden and Sorensen 2005). Mobile RAs assist in-store consumers in reaching a product choice by providing resources for comparison shopping (Youll, Morris and Maes 2001). The in-store context is a significant and prominent factor in the realization of mobile RA effects (Lee and Benbasat 2003; Lee and Benbasat 2004). The in-store context fundamentally changes the way consumers interact with mobile devices and reach a decision (Lee and Benbasat 2003; Lee and Benbasat 2004). Specifically, in-store consumers must conduct two tasks simultaneously in order to shop with the help of mobile RAs: (1) they must examine products displayed in the store and (2) they must utilize the guidance provided by the mobile RAs. Therefore, the compatibility between the way the store displays the products and the way RAs guides consumers' decision making is an important factor in predicting consumer decision performance (Slovic, Griffin, Tversky and Hogarth 1990). If the in-store display and the RA's guidance direction are compatible, a consumer can utilize the RA's guidance instantly. When there is incompatibility, the consumer must convert the format in which information is collected from the store into the manner in which the RA guides his/her decision making, a process which increases his/her mental effort and the possibility of making errors (Slovic et al. 1990). In order to examine the effects of compatibility, I developed two RAs that guide consumers' purchase choices in different ways, attribute-driven and alternative-driven manners, respectively. The attribute-driven RA is a common type of RA developed for stationary computers: It presents a list of attributes and asks consumers to specify their preferences for attributes. The alternative-driven RA, on the other hand, presents a list of 5 product alternatives and supports consumers in comparing products side by side. There is a clear limit to displaying complex products sorted by every attribute on static product shelves in a store, and consequently product alternatives are often displayed randomly by alternative or less systematically by one or two attributes (e.g., price or brand). When a consumer wants to review the details of a product s/he has found interesting in the store, s/he can check the product information instantly by clicking the product name on the list provided by the alternative-driven RA. In contrast, a consumer using the attribute-driven RA and wishing to check product details must first specify his/her preferences for attributes. In other words, the consumer must convert his/her way of acquiring information (i.e., by alternative) into another manner (i.e., by attribute) in order to use the attribute-driven RA. Therefore, the attribute-driven RA is less compatible than the alternative-driven RA for in-store purchase decisions, and comparisons of the attribute-and alternative-driven RAs will reveal the effects of compatibility between the RA's guidance direction and the store's product displays on consumers' decision processes and outcomes in a retail store. In addition, in Study 2 , I employed a control group and contrasted it with the RA groups to investigate whether the use of mobile RAs improves consumers' in-store decision-making compared to non-use. The control group was necessary because there was no previous research that empirically showed the advantages and disadvantages of using mobile RAs, compared to non-use. Thus, this dissertation provides a first step that may serve as a cornerstone upon which I compare RA-AL and RA-AT. 1.1 R E S E A R C H O B J E C T I V E S A N D Q U E S T I O N S The research questions addressed by the two studies in this dissertation are as follows: 6 • Does the R A that employs a compensatory strategy, as compared to the R A that employs a non-compensatory strategy, highlight attribute conflicts and thereby increase consumers' perceived effort, and decrease perceived control over manipulations of the R A and perceived quality of recommendations, and finally their intentions to use the R A ? • Does task emotionality moderate the effects of attribute conflicts on consumers' perceptions of control, effort, recommendation quality, and usage intentions? • Does a compatible R A , as compared to an incompatible R A , decrease in-store consumers' perceived effort and increase decision accuracy, perceived control, and intentions to use the R A ? • Does use of mobile R A s , as compared to R A non-use, decrease in-store consumers' perceived effort, increase decision accuracy, and increase intentions to return to the store where consumers shopped with the RAs? 1.2 M E T H O D O L O G Y AND M A J O R FINDINGS To investigate these research questions, two controlled laboratory experiments were conducted. Experiment 1 addressed the importance of attribute conflicts in influencing consumers' perceptions and acceptance of R A s . In addition, the moderating effects of task emotionality on the above relationship were investigated. Experiment 2 examined the influence of compatibility between the store's product displays and the R A ' s guidance on consumers' perceptions and acceptance of R A s . In addition, the advantages of mobile R A use for in-store purchases compared to R A non-use were addressed. The major findings of the two experiments are summarized as follows: (1) The R A that employs a compensatory strategy, as compared to the R A that employs a non-compensatory strategy, highlights attribute conflicts, and 7 thereby negatively influences consumers' perceived control, and recommendation quality, and intentions to use the RA. (2) Task emotionality moderates the above relationship. Consumers are more motivated to cope with high-emotion tasks, which are associated with severe negative consequences, than with low-emotion tasks, which they associate with little harm to themselves. As a result, the gap between the R A employing a compensatory strategy and the R A employing a non-compensatory strategy becomes augmented for high-emotion tasks as compared to low-emotion tasks. (3) Compatibility between the store's display and the RA ' s guidance direction increased both consumers' perceived control and decision accuracy. However, compatibility did not influence consumers' perceived effort or intentions to use the RAs. (4) Use of mobile RAs, as compared to R A non-use, reduced in-store consumers' perceived effort and increased their intentions to return to the retail store where they had shopped with the RA. 1.3 STRUCTURE OF THE DISSERTATION The remainder of the dissertation is structured as follows. Chapter 2 begins with a review of decision strategies associated with the four RAs employed in Studies 1 and 2. As the four RAs are built based upon one (or more) strategies, a review of the strategies demonstrates the characteristics of the four RAs clearly. Next, I review the existing literature about factors influencing acceptance of RAs; I thereby establish how important it is to examine attribute conflicts in investigating consumers' acceptance of RAs. I demonstrate the lack of prior research on compatibility, and illustrate the importance of investigating this issue. In Chapter 3,1 introduce the first experiment, in which the effects of attribute conflicts on accuracy, effort, control, and the influence of all of these on 8 consumers' intentions to use RAs, are investigated. In Chapter 4, I describe the second experiment, in which the effects of using mobile RAs on consumers' purchase choices, and the effects of compatibility on accuracy, effort, control, and intentions to use the RAs, are examined. Finally, in Chapter 5,1 summarize the findings from both studies, outline the major contributions made by this dissertation, and provide suggestions for future research. 9 C H A P T E R 2: L I T E R A T U R E R E V I E W In this chapter, I introduce the six decision strategies associated with Studies 1 and 2 and describe the differences among them; then, I present the constructive decision-making theory to explain how these differences influence consumers' acceptance of RAs that embed the strategies. Next, I outline previous research on the influence of attribute conflicts on consumers' acceptance of RAs, which is the theme of Study 1. This outline manifests conflicting results from a few existing studies on attribute conflicts and suggests a solution to resolve these conflicts. Lastly, I review previous research on mobile RAs used for in-store purchase decisions, the topic investigated in Study 2. The review demonstrates the lack of empirical studies on this topic. I thereby highlight the need for an investigation into this topic. 2.1 DECISION STRATEGIES This section describes the decision strategies associated with the RAs of interest to both Studies 1 and 2 based upon Payne et al. (1993). Consumers can apply 12 different decision strategies to multi-alternative/multi-attribute preferential choice problems that involve choosing one out of a number of alternatives described by a common set of attributes (Svenson 1979). Of these, the two RAs of interest to Study 1 employed the Weighted-Additive (WAD) strategy and the Elimination-By-Aspects (EBA) strategy, respectively (hence, RA-WAD and RA-EBA) . WAD and E B A are at opposite extremes, among other things, in terms of the extent to which they highlight attribute conflicts, the main focus of Study 1 (Bettman, Johnson, Luce and Payne 1993; Bettman et al. 1998; Payne et al. 1993; Widing and Talarzyk 1993). In addition to these two strategies, ADditive-DIFference (ADDIF), Majority of Confirming Dimensions (MCD), EQual Weight (EQW), and Satisficing (SAT) strategies 10 are related to the RAs in Study 2. Use of these strategies can be invoked in conjunction with the two RAs that employ distinct guidance directions: alternative-driven vs. attribute-driven approaches (hence, R A - A L and RA-AT). Unlike RA-WAD and R A - E B A employed in Study 1, R A - A L and RA-AT are loosely coupled with decision strategies. In other words, RA-WAD and R A - E B A both compel consumers to use a particular dedicated strategy, i.e., WAD and E B A , respectively. Conversely, R A - A L and RA-AT offer consumers interface features with which they can employ strategies at their own discretion. In other words, R A - A L users are given choices of ADDIF, M C D , and SAT, and EQW, while RA-AT users are given choices of WAD and E B A . The detailed reasons for why and how these particular strategies are associated with R A - A L and RA-AT are provided in the section 4.2.2. The six strategies are described in detail below. The descriptions below are based upon Payne et al. (1993). WAD is based on the evaluation of one alternative at a time along all relevant attributes. Each attribute is assigned a weight. A score for each alternative is determined by adding the product of the attribute value and the weight. Use of WAD hence compels decision makers to process all the information about each alternative and to embark upon cumbersome computations, such as multiplications and summations. EBA is based on an elimination heuristics and allows decision makers to exclude alternatives without processing all the relevant information about them, which simplifies decision making. E B A compares attribute values against user-specified threshold levels across all alternatives. Those alternatives with attribute levels below the threshold levels are eliminated. Such E B A processes continue until one or a few alternatives remain. In the ADDIF strategy, a pair of alternatives is compared for each attribute; the difference between the two in each attribute is calculated (Tversky 1969). Then each difference is weighted, and the weighted differences are added up over all attributes to 11 obtain an overall relative evaluation score of the two alternatives. The winning alternative is retained while the other is eliminated. This process continues until the final winning alternative is identified (Todd and Benbasat 1992). ADDIF is equivalent to WAD in that both are compensatory strategies that produce identical preference orderings under some conditions (Payne et al. 1993). Unlike WAD, ADDIF asks users to assess the importance weights of attributes based on pair-wise comparisons of alternatives. MCD simplifies ADDIF by coding the attribute differences in a dichotomous manner (i.e., whether an alternative A is better [or worse] than B in terms of the alternative) without weighting the differences. Hence, only the direction of the difference, but not its relative size, is counted; the alternative with more of winning (better) attribute values is maintained temporarily. As is the case of ADDIF, the chosen alternative is then compared to the next alternative and this pair-wise comparison is repeated until the final winning alternative is selected. E Q W strategy simplifies WAD by ignoring the relative importance weights of the attributes and assumes the same weight for each attribute. EQW, therefore, is considered a special case of WAD, in which users consider each of the attributes equally important. However, EQW has been advocated as a highly accurate simplification of the decision making process and does not lower the accuracy of WAD in many cases (Einhorn and Hogarth 1978). Users of the RAs employing EQW strategy reported the same level of accuracy and effectiveness as those of the RAs employing WAD (Lee, Liu and Yeo 2002; Song et al. 2007). With S A T strategy, alternatives are evaluated one at a time, in the order they are listed in the alternative list. The value of each attribute of an alternative is compared to a predefined threshold level. An alternative that has any one of attribute values below the threshold is eliminated. The first alternative whose attribute values meet all the thresholds 12 is selected. If none of the alternatives satisfies all the thresholds, then the thresholds are lowered and this process is repeated, or an alternative is chosen randomly. Therefore, the use of this strategy is invoked when a decision maker goes through a list of alternatives and makes a quick judgment on whether or not an alternative on the list is good enough to be chosen. 2.1.1 Properties of Decision Strategies I contrast the six strategies in terms of some of general properties, relevant to the two studies in this dissertation, outlined by Payne et al (1993). According to Payne et al. (1993), strategies have emerged from different disciplines and are therefore described using different conventions. Therefore, making comparisons using these general properties helps one to recognize commonalities and differences among these strategies. Table 2.1 outlines the general properties of the six decision strategies. - INSERT [Table 2.1] H E R E -Compensatory versus non-compensatory. Depending upon "the ability of a good value on one attribute to make up for bad values on other attributes" decision strategies are categorized as compensatory or non-compensatory. E B A and SAT are non-compensatory because an inferior value on one attribute will prevent an alternative from being chosen, regardless of how superior it is on other attributes. WAD, ADDIF, and EQ are compensatory because good values on one attribute can compensate for bad values on others. M C D rule is partially compensatory in that the total number of better attributes for an alternative counts, but the magnitude. This distinction between compensatory and non-compensatory strategies is associated with how a strategy deals with attribute conflict. An attribute conflict arises when the favourable values of some attributes are correlated with the unfavourable values 13 of other attributes (Luce et al. 1997). In an example of rental apartments, proximity to public transportations tends to be associated with a higher monthly rent. Compensatory rules reveal conflict, whereas non-compensatory rules hide it. For instance, WAD, one of compensatory strategies, requires users to confront conflicts and to make explicit tradeoffs (Luce et al. 1997; Luce et al. 1999). While reviewing the attributes of an alternative, a WAD user is alerted to any conflicts and to upcoming tradeoffs s/he needs to make. In contrast, E B A , one of non-compensatory strategies, does not require users to assess attribute conflicts or to confront inevitable tradeoffs (Davis and Kottemann 1994). An E B A user eliminates an alternative before examining all its attribute values, thus bypassing an assessment of attribute conflicts (Hogarth 1987). Amount of processing (Information ignored). This concerns whether strategies explicitly ignore possibly relevant information in conducting a decision task, and hence decreases the amount of information processed, or process all relevant information. WAD and ADDIF take into account all relevant information, while EQW ignores relative importance weights (as it assumes equal weights for each attribute); E B A and SAT eliminate alternatives below threshold levels; M C D does not weight differences between the attributes of two alternatives being compared. Consistent versus selective processing. This concerns "the extent to which the amount of processing is consistent or selective across alternatives or attributes" (Payne et al. 1993, page 30). That is, it questions whether or not the amount of information examined for each alternative or attribute varies. Consistent processing in general means that all information for every alternative and attribute is examined. More consistent processing across alternatives is a necessary condition for an accurate compensatory strategy. Selective processing implies a strategy of eliminating alternatives or attributes by using only partial information, without reckoning whether additional information may alter the final choice. 14 2.1.2 Comparisons between W A D and E B A As mentioned earlier, the RAs designed for Study 1 employed WAD and E B A strategies, because they contrast with each other in terms of the three prominent predictors of consumers' strategy choice: (1) the effort required of decision makers to make a choice, (2) the accuracy of resulting outcomes, and (3) the extent to which attribute conflicts are highlighted (Bettman et al. 1993; Bettman et al. 1998; Payne et al. 1993; Widing and Talarzyk 1993). As Table 2.1 shows, WAD does not ignore information and processes information consistently, while E B A ignores information about eliminated alternatives and selectively processes information about only the alternatives that remain. As a result, WAD requires more effort from decision makers than does E B A . Using the WAD strategy yields more accurate choices than using E B A , since WAD employs compensatory processing whereas E B A involves non-compensatory processing. Lastly, WAD involves a form of compensatory processing, which is known to reveal and confront attribute conflicts. Consequently, WAD manifests attribute conflicts more substantially than E B A . 2.1.3 Comparisons between ADDIF , E Q W , M C D , SAT and W A D , E B A Study 2 employs two RAs: R A - A L (alternative-driven RA) and RA-AT (attribute-driven RA). "Attribute-driven" and "alternative-driven" indicate the sequences in which the RAs guide consumers' decision making. R A - A L enables the use of ADDIF, EQW, M C D , and SAT, while RA-AT allows the use of WAD and E B A (see Section 4.2.2 that describes the two RAs and explains why and how the two RAs are associated with these particular strategies.). ADDIF and WAD are the two more compensatory strategies, and thus generate highly accurate results while involving heavy processing of information on the decision-maker's part. The rest - EQW, M C D , SAT and E B A - are less accurate given that they ignore some information. Nevertheless, EQW and M C D are compensatory and 15 process information consistently, whereas SAT and E B A are non-compensatory and process information selectively. 2.2 CONSTRUCTIVE DECISION MAKING THEORY Consumer decision-making is constructed on the fly influenced by a number of contextual factors, such as, the goals they want to achieve in a particular decision context, characteristics of the decision task, display and representations of the task, and so on. (Bettman et al. 1998). This notion contrasts a traditionally accepted view of consumers as rational decision makers who apply well-established preferences towards products, strategies and approaches regardless of decision tasks. Rather than applying the optimal strategy invariantly to any decision contexts, consumers develop a hierarchy of goals, on the fly, that they want to achieve in a particular situation, and choose an approach that best supports the attainment of the goals (Bettman et al. 1998). The goals that capture many of the most important motivational aspects relevant to decision-making are: maximizing accuracy of the decision, minimizing cognitive effort to make a decision, and minimizing the experience of negative emotions arising from attribute conflicts (Bettman etal. 1998). 2.3 EFFECTS OF ATTRIBUTE CONFLICTS ON CONSUMERS' PERCEPTIONS AND ACCEPTANCE OF RECOMMENDATION AGENTS This section reviews previous research on the influence of attribute conflicts on consumers' acceptance of RAs, and thereby illustrates both the lack of existing relevant empirical research, and the conflicting propositions and results reported by previous research. Based upon such conflicting findings, this section derives two research objectives, which are to empirically examine: (1) the overall influence of attribute conflicts on users' acceptance of RAs, and (2) the moderating effects of task-emotionality on users' R A acceptance. This is a generally under-studied area, and only three empirical studies have been 16 found, as shown in Table 2.22. These three studies adhere to two opposite tenets. Widing and Talarzyk (1993) and Fasolo et al. (2005) claim that individuals prefer RA-WAD precisely because it presents attribute conflicts, on the grounds that individuals actively attempt to resolve attribute conflicts to reach consistent and reliable choices (Delquie 2003; Hsee, Loewenstein, Blount and Bazerman 1999). On the contrary, Aloysius et al. (2006) contend that individuals prefer a DSS that hides attribute conflicts because attribute conflicts invoke decisional stress and consumers will try to avoid confronting attribute conflicts so as to minimize the experience of decisional stress (Aloysius et al. 2006; Davis and Kottemann 1994). The three studies are discussed in greater detail below. - INSERT [Table 2.2] H E R E -Widing and Talarzyk (1993) compared two Decision Support Systems (DSS) - a cutoff format DSS (DSS-EBA) and a linear format DSS (DSS-WAD) - and a plain matrix of attributes by alternatives for three tasks, varying the degree to which attributes are negatively correlated. DSS-EBA and DSS-WAD are equivalent to R A - E B A and R A -WAD except that the DSS are not web-based while RAs are web-based. They argue that DSS-WAD yields the most accurate results, followed by DSS-EBA and the matrix, as negative correlations among attributes increase. DSS-WAD employs the compensatory strategy which is insensitive to any contextual factors (i.e., attribute conflicts), while the accuracy of DSS-EBA decreases as attribute conflicts increase, since the choice of a desirable level for an attribute results in the elimination of alternatives with desirable values for the other attributes. Nevertheless, they note that both DSS-WAD and DSS-E B A users make more accurate choices than matrix users, that is, DSS in general enhance decision quality. Widing and Talarzyk measured decision accuracy by asking users about their intentions to switch to another alternative when given the chance to do so. 2 Refer to Xiao and Benbasat's (2007) review paper on previous research on RAs. 17 In Widing and Talarzyk's laboratory study, participants were asked to choose word-processing software which included 30 brands rated on six attributes, such as ease of start-up, ease of learning, ease of use, error handling, performance, and versatility. The results largely support their hypotheses. Significantly fewer DSS-WAD users reported switching intentions than did DSS-EBA and matrix users; this tendency became more manifest as the attribute conflicts increased. However, the participants' subjective perceptions of the DSS did not entirely match their switching intentions. There were significant differences between DSS users and matrix users in terms of perceived accuracy, decision difficulty, time-saving, and satisfaction with the aids they used. DSS-WAD and DSS-EBA users, however, did not show significant differences in any of these subjective aspects, regardless of the magnitude of attribute conflicts. Fasolo et al.'s (2005) claim parallels Widing and Talarzyk's (1993). They compared two RAs - RA-WAD and R A - E B A . They argued that when attributes are positively correlated, people tend to use R A - E B A , since an option that is good on one attribute is also good on other attributes, such that setting a threshold on more than one attribute does not exclude options prematurely. With negative inter-attribute correlations, however, users tend to rely on a compensatory strategy, because users' natural response is to confront trade-offs and to process information by option. They hypothesized that users should find decision making easier when faced positive rather than negative correlations, and that RA-WAD makes decision processing easier and more effective than R A - E B A , especially in the presence of negative inter-attribute correlations. The participants in Fasolo et al.'s study were asked to choose one out of 40 digital cameras described with seven attributes, such as delay between shots, image capacity, LCD display, light sensitivity, zoom, resolution, and weight. The results supported most of their hypotheses. Users' effort (measured by the number of clicks) increased when attributes were negatively correlated, as opposed to when they were positively correlated. 18 RA-WAD users chose more non-dominated options (i.e., objectively better options in every aspect) than R A - E B A users, and this difference became larger as the attributes were negatively correlated. Simultaneously, R A - E B A users were less satisfied with, and less confident in, the chosen options, found the decision task more difficult, and were less satisfied with the RA, than RA-WAD users, especially when attributes were negatively correlated. Aloysius et al.'s (2006) arguments and findings oppose those in the above two studies. They employed two DSS, each of which used a distinct preference-elicitation method: one involving pair-wise comparisons of attributes (similar to RA-WAD) and the other involving an absolute measure of attributes (although there is no equivalent to it in the current study, this is similar to R A - E B A in the sense that this preference-elicitation method is based upon within-attribute choices). Aloysius et al. posit that users preferred the DSS that employed an absolute measure to the DSS that involved pair-wise comparisons. This is because pair-wise comparisons require users to assess attribute conflicts and to make explicit tradeoff judgments among conflicting attributes, thereby increasing decisional stress and negatively influencing users' perceptions of DSS. In addition, Aloysius et al. argue that attribute conflicts highlighted by the pair-wise comparison DSS negatively influenced users' perceived accuracy and increased the perceived effort involved in using DSS. Aloysius et al.'s laboratory experiments required participants to choose one out of three job offers described by four attributes, namely starting salary, quality of life, interest in work, and proximity to family and friends. The results show that users of the pair-wise comparison DSS reported higher decision conflicts, lower accuracy, higher effort, and consequently, lower preferences than those of the absolute measure DSS. These three studies clearly illustrate two very contrasting perspectives to explain 19 the roles of attribute conflicts in users' perceptions and acceptance of RAs. In addition, the decision tasks employed by the three studies varied significantly in terms of the severity of negative consequences brought about by suboptimal decisions. A bad career choice in Aloysius et al. (2006)'s study had a higher chance of bringing more negative consequences because the attributes associated with job offers, such as quality of life and interest in work, are more difficult to regain once lost. Conversely, an error in choosing the right software package and digital camera can easily be corrected with a replacement r at minimal or no cost (assuming that most retail stores have a free replacement policy within 2 weeks from the date of purchase). According to the theory of constructive decision making, the decision task is a. major influence on users' decision-making processes and outcomes (Payne et al. 1993; Slovic 1972). The constructive-preferences perspective claims (a) that expressions of preference are generally constructed at the time at which the valuation of an object is required and (b) that this construction process is shaped by the interaction between the properties of the human information-processing system and the properties of the decision task. In a similar vein, Slovic (1995) noted that preferences appear to be remarkably adaptive, sensitive to the way a choice problem is described and framed (Slovic 1995). The existence of rich empirical evidence that supports the idea that the particular characteristics of the decision environment play a central role in individuals' decision-making, suggests that the conflicting results found in previous research may be attributed to the characteristics of the choice tasks employed. It should be recalled that attribute conflicts signal the prospect that certain attributes must be traded off and that the goals associated with the attribute will be lost. When such losses are linked to highly negative consequences (such as low life quality in Aloysius et al.'s study (2006)) and the best course of action is unclear, consumers are more likely to feel decisional stress and thus attempt to avoid confronting the attribute conflicts (Luce et al. 2001; Tetlock, Kristel, 20 Elson and Green 2000). Consequently, consumers will negatively evaluate an R A that forces them to confront the emotion-laden attribute conflicts but fails to provide a clear way to resolve those conflicts. However, when such losses affect consumers' well-being to a lesser degree (such as the versatility of word-processing software in Widing and Talarzyk's 1993 study), those consumers are not likely to feel intense stress while trading off certain attributes. Therefore, their perceptions of the conflict-confronting R A may be influenced less negatively or even positively if they believe that the confrontations lead to better results. In summary, the review of previous research reveals a need for thorough empirical research that investigates the influence of attribute conflicts on users' perceptions and adoptions of RAs. In addition, the conflicting results reported in previous studies suggest that further examination is required into the moderating effects of specific tasks, such as tasks involving highly negative consequences. Study 1 attempts to respond to these requirements. Specifically, in Study 1, R A -WAD and R A - E B A are compared in terms of the three prominent factors influencing users' acceptance of RAs: (1) perceived effort, (2) perceived quality of recommendations, and (3) perceived control (i.e., the degree to which RAs make users believe that they can resolve attribute conflicts). By examining these three factors, I will illustrate how attribute conflicts influence consumers' intentions to use RAs negatively. Additionally, I investigate the moderating effect of task emotionality on such relationships. Task emotionality is defined as the degree to which a decision task is perceived to bring severe negative consequences to the well-being of people decision makers care about (Luce et al. 1997; Luce et al. 1999). In order to observe the moderating effect of task emotionality, I compare users' perceptions and acceptance of RA-WAD and R A - E B A across high- and low-emotion tasks. By investigating the moderating effect of task emotionality, one can identify the reasons for the equivocal results obtained from the prior studies. 21 2.4 M O B I L E R E C O M M E N D A T I O N A G E N T S F O R I N - S T O R E D E C I S I O N M A K I N G This section outlines previous research on mobile RAs used for in-store purchase choices in Study 2. In general, there is a lack of research in this field. Among the few existing studies, most of them focus on the technical development of mobile RAs, while none of the remaining empirical studies examined the influence of compatibility between RAs' guidance directions and stores' product displays on consumers' acceptance of RAs. As in the previous section, I first provide a summary of existing studies, and then explain the rationale for conducting Study 2. Miller et al. (2003) examined the challenges associated with building a recommendation system for four wireless interfaces: (1) A P D A using the AvantGo service (file-synchronization software that transfers Web pages from a PC to a PDA), (2) a web-enabled cell phone browser, (3) a voice interface over the telephone, and (4) a wireless PDA. They built "MovieLens Unplugged," a companion service to the desktop-platform "MovieLens" website, which provides recommendations to consumers in a movie-rental store. They conducted a field experiment in which real consumers used the interfaces while in actual movie rental stores, and a laboratory experiment, in which participants conducted two tasks provided by the researchers. Miller et al. did not provide statistical analyses but instead interviewed the participants. Most of the participants indicated positive responses towards mobile recommendation systems and clearly saw the benefits of receiving recommendations at the point of decision (Miller, Albert, Lam, Konstan and Riedl 2003). The Impulse research project at the MIT Media Laboratory explored a location-based mobile recommender system at point of purchase (Youll et al. 2001). This is a location-based system that provides a dynamic list of products or stores nearby. The MIT researchers advocated the importance and benefits of the in-store assistance of RAs and 22 contended that such assistance enables consumers to negotiate with merchants more strongly because it allows them to access product reviews and alternative offerings. They did not conduct any empirical tests involving human subjects. In a series of studies, Van der Heijden and his colleagues investigated consumers' product choices with the help of a mobile R A that employed the WAD strategy in a retail store created in a laboratory (Van der Heijden 2005; Van der Heijden and Sorensen 2002; Van der Heijden and Sorensen 2005). To the best of my knowledge, to date these have been the only empirical investigations into consumers' decision making with mobile RAs, employing decision strategies, in a retail-store setting. In all three studies, the same type of R A (i.e., the R A with WAD support) was developed and compared with a plain mobile website providing product information only, that is, without an RA. Nevertheless, the R A with WAD support did not fully computerize or implement WAD. Instead of the R A eliciting consumers' preferences for products, human research assistants collected participants' preferences on paper-based questionnaires, calculated an overall score for each alternative, and entered the scores into the R A manually. Then, participants entered the store and began shopping with the RA. During the shopping session, the R A with WAD support marked alternatives with higher scores with a light-coloured "attractiveness cue," and alternatives with lower scores with a darker-coloured cue. Van der Heijden and Sorenson (2002) and Van der Heijden (2005) compared the R A and the plain mobile website in terms of consumers' subjective confidence in decisions and objective accuracy of decisions, which were measured by the number of non-dominated/dominated products included in the consideration set (the group of products that remained after the initial screening). In addition, they investigated the moderating effect of task complexity - i.e., the number of alternatives included in the 23 study. The results showed that the R A with WAD support significantly increased the number of non-dominated alternatives included in consideration sets, but did not increase users' subjective decision confidence. Task complexity significantly moderated the effects of WAD support on decision confidence: WAD support increased users' confidence for a complicated task more than for a less complicated task (Van der Heijden and Sorensen 2002). Van der Heijden and Sorensen (2005) contrasted the R A and the plain mobile website in terms of participants' utilitarian and hedonic ratings. They found no significant difference in the ratings between the R A and the plain website. The above summary of previous studies shows that researchers have recognized the significant role of mobile RAs in helping consumers in a retail store. Consumers often lack the resources to conduct comparison shopping in retail stores, and mobile RAs could fulfil this immediate need (Youll et al. 2001). Despite the consensus on the potential value of such RAs, there has been little behavioural research in this area, except Van der Heijden and his colleagues' three studies. As the very first researchers to delve into empirical studies in this field, they provided a cornerstone for this topic. Expanding from their studies that compared RA-WAD with a plain website in terms of decision accuracy, I compare two RAs that employ two distinct guidance directions that are more or less compatible with the store's product displays. I investigate whether or not compatibility between RA's guidance directions and in-store displays influence consumers' decision accuracy, perceived effort, control, and intentions to use RAs. Simultaneously, in Study 2, I compare R A users with R A non-users in terms of decision accuracy, perceived effort, and intentions to return to the store where they shopped with the RAs. 24 2.5 T A B L E S Table 2.1 Properties of the Six Decision Strategies (Payne et al. 1993, page 32) Compensatory (C) versus non-compensatory (N) Information Ignored? (Y orN) Consistent (C) versus Selective (S) WAD C N C ADDIF C N C E Q W C Y C E B A N Y S M C D C Y C SAT N Y S 25 Table 2.2 Summary of Previous Research on Attribute Conflicts Perspective Authors Main Hypotheses Results Decision Tasks Employed • Individuals attempt to resolve attribute conflicts actively to reach consistent and reliable choices. • A s a result, individuals prefer an RA that precisely presents attribute conflicts. Widing and Talarzyk(1993) Objective accuracy of decisions yielded by use of DSS-WAD will be higher than that of D S S - E B A . Supported • To choose one out of 30 word-processing software brands • Six attributes: ease of start-up, ease of learning, ease of use, error handling, performance, and versatility Subjective perceptions of D S S -WAD, including perceived accuracy, decision difficulty, time-saving and satisfaction with the DSS will be higher than those of D S S - E B A . Not Supported Fasolo et al. (2005) Objective accuracy of decisions yielded by use of RA-WAD will be higher than that of RA-EBA. Supported • To choose one out of 40 digital cameras • Seven attributes: delay between shots, image capacity, L C D display, light sensitivity, zoom, resolution, and weight. Subjective perceptions of RAs-WAD, including satisfaction with, and confidence in, the final choice, decision difficulty, and satisfaction with the RA will be higher than those of RA-EBA. Supported • Individuals feel decisional stress dealing with attribute conflicts and hence try to escape from confronting the conflicts to minimize the experience of decisional stress. • As a result, individuals prefer a D S S that hides attribute conflicts. Aloysius et al. (2006) Subjective perceptions of a DSS emphasizing attribute conflicts, including perceived accuracy, effort-expenditure, and preferences, will be lower than those of a D S S that hides the conflicts. Supported • To choose one out of three job offers • Four attributes: starting salary, quality of life, interest in work, and proximity to family and friends. 3 Only the main hypotheses that are related to the comparisons between R A types are summarized due to space constraints. C H A P T E R 3: E F F E C T S O F A T T R I B U T E C O N F L I C T S O N C O N S U M E R S ' PERCEPTIONS AND A C C E P T A N C E O F R E C O M M E N D A T I O N A G E N T S 3.1. I N T R O D U C T I O N Recommendation Agents (RAs), web-based intelligent recommendation systems, advise consumers what to buy based upon the needs expressed by the consumers (Wang and Benbasat 2007). RAs reduce information overload and facilitate a more accurate choice of products (Maes et al. 1999; O'Keefe and McEachern 1998). The role of different RAs has previously been examined from the perspective of the effort-accuracy framework, specifically how RAs save effort and increase the accuracy of decision-making (Haubl and Trifts 2000; Xiao and Benbasat 2007). In the current study, I expand the effort-accuracy perspective to investigate the degree to which different types of RAs support users in coping with negative correlations among product attributes (i.e., attribute conflicts). An attribute conflict occurs when the favourable values of some attributes are correlated with the unfavourable values of other attributes (Luce et al. 1997). For instance, advanced safety features of an automobile tend to be associated with a higher price. Consumers dealing with attribute conflicts are compelled to trade off between attributes and to accept less of the one for more of the other. As a result, they often experience decisional stress, as their attainment of an important goal becomes threatened or actually blocked by their chosen course of action and they begin to anticipate blame for, or regret about, the unattained goal (Lazarus 1991). Appraising the lost goal, consumers become concerned that forgone alternatives might be better than the chosen one and that they may come to regret their choice later (Zeelenberg 1999). Despite their importance in users' decision-making, attribute conflicts have been 27 understudied by current R A and DSS literature. A handful of existing studies on this issue have provided conflicting results and arguments, failing to reach a consensus. Furthermore, the current designs of many RAs focus on increasing decision accuracy and reducing users' effort, while often unintentionally highlighting attribute conflicts which are the prominent cause of decisional stress (Hogarth 1987). The decisional stress consumers experience while using RAs inhibits them from adopting RAs (Davis and Kottemann 1994). By further examining attribute conflicts, I will examine one of the prominent reasons why some consumers choose not to use RAs that allegedly enhance their decision-making and why they prefer simple RAs when they do use them (Davis and Kottemann 1994). This research not only yields suggestions for better R A designs, but also contributes to the advancement of R A and DSS theories, which concerns themselves with how IS facilitate consumer decision-making and hence are accepted by consumers. Thus, my first goal in this study is to examine whether or not, and in what ways, attribute conflicts negatively influence consumers' perceptions of, and intentions to use, RAs. To achieve this goal, I compare two types of RAs: RA-WAD (RA applying Weighted ADditive strategy) and R A - E B A (RA applying Elimination-By-Aspect strategy), which represent the opposite extremes in presenting attribute conflicts to users. RA-WAD automates the WAD strategy which selects the alternative with the highest utility score (i.e., the total of a user's assigned utility scores for each attribute multiplied by the standardized values of the corresponding attributes). In order to implement the WAD strategy, RA-WAD requires users to distribute a total of 100-point importance weights to the attributes of a product. This means that a consumer must assign fewer weights to some attributes in order to assign more to others. As a result, the consumer becomes more aware of attribute conflicts and confronts the prospect that tradeoffs among different attributes must be made. R A - E B A automates the E B A rule that 28 eliminates alternatives that do not meet user-specified thresholds. R A - E B A , therefore, allows the user to avoid facing explicit attribute conflicts, since it simply asks users to choose the least acceptable levels of each attribute, rather than requiring across-attribute comparisons that make users more acutely conscious of existent attribute conflicts. I compare the two RAs in terms of four dependent variables: (a) consumers' perceived control over manipulations of RAs to specify their preferences, (b) consumers' perceived effort required to use RAs to make decisions, (c) consumers' perceived quality of recommendations provided by RAs, and finally (d) consumers' intentions to use RAs. The first three dependent variables represent the three most important motivations in a user's choice of decision aids: minimizing the confrontations with attribute conflicts during decision-making processes, minimizing effort required to make a decision, and maximizing accuracy, respectively (Bettman et al. 1998). By examining these three variables, I illustrate the ways in which attribute conflicts influence usage intentions negatively. Specifically, I argue that, due to the conflict-highlighting nature of RA-WAD, RA-WAD users will perceive less control and more effort, evaluate quality of recommendations less favourably, and show lower usage intentions than R A - E B A users. My second goal in this study is to investigate the moderating effect of task emotionality on the influence of attribute conflicts on users' perceptions and usage intentions of RAs. Task emotionality is defined as the degree to which a decision task is perceived to bring severe negative consequences to the well-being of people decision-makers care about (Luce et al. 1997; Luce et al. 1999). Because high-emotion tasks involve attribute conflicts which lead to more harmful compromises for consumers, the use of RA-WAD, which highlights attribute conflicts, becomes a more unpleasant experience for high-emotion tasks than for low-emotion tasks. In contrast, R A - E B A enables consumers to avoid confrontations with attribute conflicts for both high- and low-emotion tasks; as a result, their perceptions of R A - E B A are rather consistent across both 29 high- and low-emotion tasks. Thus, the differences between RA-WAD and R A - E B A are expected to be more pronounced for high-emotion tasks than for low-emotion tasks. It is important and necessary to examine the moderating effects of task emotionality, in order to counter the alternative hypothesis that factors other than attribute conflicts cause users' perceptions of RA-WAD to differ from their perceptions of RA-E B A . Task emotionality determines the degree to which attribute conflicts influence consumers' decision-making without altering other cognitive factors associated with the decision problem (e.g., the numbers of alternatives and/or attributes, or the degree to which attributes are negatively correlated) (Luce et al. 1997; Luce et al. 1999). Therefore, if the differences between RA-WAD and R A - E B A are greater for a high-emotion task than for a low-emotion task, one can conclude that this difference is caused solely by the attribute conflicts. In order to achieve the two goals of this study (to examine how attribute conflicts influence consumers' perceptions of, and intentions to use, RAs and investigate the moderating effect of task emotionality on this process), I conducted a laboratory experiment using a 2 x 2 factorial design with two between-subject factors: R A types and task emotionality. The two types of RAs were RA-WAD and R A - E B A ; task emotionality included high- and low-emotion levels. Following previous studies, I altered task emotionality by (1) varying the vividness of the described negative decision consequences by providing more cues illustrating potential consequences to the high-emotion group, and (2) varying the reference points that describe the choice as an improvement or a downgrade from the product currently owned (Luce et al. 1997; Luce etal. 1999). The remainder of this chapter is organized as follows. In Section 3.2,1 describe the independent variables (i.e., the two types of RAs), and the dependent variables (i.e., 30 perceived control, recommendation quality, perceived effort, and usage intentions) and develop hypotheses about the overall influence of the attribute-conflicts portrayed by the two RAs on the dependent variables. In addition, I illustrate the moderating variable (i.e., task emotionality) and the ways in which it alters such relationships. In Section 3.3, I present research method that empirically compares RA-WAD and R A - E B A across high-and low-emotion tasks. In Section 3.4, I present the data analyses and results. I provide A N C O V A results to illustrate whether or not the hypothesized main and interaction effects are present, followed by the results of a structural equation modeling to further investigate how perceived control, quality, and effort affect usage intentions. Finally, in Section 3.5, I discuss the results of the experiment and its implications for theory and practice, and make suggestions for future research. 3.2 THEORY DEVELOPMENT 3.2.1 R A - W A D and R A - E B A RA-WAD and R A - E B A are typical computerized implementations of WAD and E B A strategies (Haubl and Trifts 2000; Widing and Talarzyk 1993). RAs consist of two major components: (1) a preference-elicitation method which collects an individual decision-maker's multi-attribute preferences with respect to a particular domain or product category, and (2) an algorithm that makes recommendations in the form of a sorted list of alternatives based on the individual's preference structure (Haubl and Murray 2003; Haubl and Trifts 2000). Below, I describe how the two key elements of RA-WAD and R A - E B A implement the underlying strategies, thereby influencing users' perceived control, recommendation quality, effort, and usage intentions. RA-WAD asks users to indicate how important they personally consider each of the attributes to be on a 100-point constant-sum scale. Multiplying these subjective attribute-importance weights by standardized scale values for the different levels of each attribute, 31 RA-WAD computes an overall utility score for each alternative and recommends the product that has the highest utility score. R A - E B A requires consumers to specify the minimum level of each attribute they will personally accept and then provides a list of products that satisfy these cutoff levels. Since these algorithms precisely reflect the underlying strategies, the accuracy achieved when using RA-WAD is higher than that achieved with R A - E B A . RA-WAD reduces the effort required for using the WAD strategy by automating the calculations required for WAD (e.g., summations and multiplication). A number of previous empirical studies have shown that RAs reduce the amount of effort required for using WAD to the level of E B A (Haubl and Trifts 2000; Todd and Benbasat 1991; Todd and Benbasat 1994; Todd and Benbasat 1996; Todd and Benbasat 1999; Widing and Talarzyk 1993). I conducted the Natural GOMS Language (NGOMSL) analyses devised by (Kieras and Helander 1997), which are particularly useful for predicting the execution time for a system prior to its implementation. The GOMS results show that users are expected to devote similar execution times to RA-WAD and R A - E B A (see Appendix 1 for detailed analyses). However, RA-WAD also highlights compelling attribute conflicts with its preference-elicitation method, which involves across-attribute comparisons (Aloysius et al. 2006). RA-WAD employs a 100-point constant-sum scale that requires the total of importance weights to be 100. This means that a consumer must assign fewer weights to one attribute in order to assign more to another. The assignment of fewer importance weights to an attribute implies the acceptance of a less favourable value of the attribute, indicating that the attribute is being traded off. Moreover, the consumer makes such tradeoffs entirely based upon his/her subjective judgment of attribute importance, without objective rationale. Therefore, the consumer becomes potentially vulnerable to blame for trading off attributes related to social norms or the well-being of others (e.g., the safety of 32 family members) to acquire other replenishable attributes (e.g., money) (Drolet and Luce 2004). R A - E B A ' s preference-elicitation method takes the form of within-attribute choices rather than tradeoff judgments. It asks consumers to specify the minimum or maximum level they require of an attribute, thereby allowing them to implicitly trade off other attributes whose values carry negative associations. For instance, the consumer can say, "I cannot afford to spend more than $5,000 on our family car, so I have eliminated cars that cost more than $5,000. The remaining options lack comprehensive safety features, but I have no choice but to select a car from among them." S/he has then compromised on safety features without admitting that s/he voluntarily traded off the safety of his/her family members in order to save money. In short, R A - E B A describes a decision task as a choice between attribute levels (e.g., $5,000, $6,000, or $7,000), rather than a tradeoff among attributes (e.g., price vs. safety features). 3.2.2 Perceived Control I argue that RA-WAD users will feel less control during the preference-specification phase than R A - E B A users, because RA-WAD's preference-elicitation method explicitly presents the fact that it is impossible to attain all the goals they want to achieve from the purchases (Luce et al. 2001), whereas R A - E B A provides users with the illusion of control (Davis and Kottemann 1994). Perceived control is defined as "the degree to which a person feels that he/she can impact one's own activities or given conditions in correspondence with higher order goals" (Frese, Ulich and Dzida 1987, p. 315). Averill (1973) defines perceived control as "whether or not an individual feels that s/he has an adequate response available to a stressor." Decisions involving attribute conflicts are the most prominent stressors because an individual's attempt to attain one attribute inevitably impinges on his/her attaining 33 another (Averill 1973; Frese, Ulich and Dzida 1987; Lazarus 1991; Van Egeren 2000). Morris and Marshall's (2004) feedback-control loop (Figure 3.1) illustrates the mechanism whereby system users experience varying levels of control: An individual sets an internal reference point that represents the ideal sets of his/her environment. Through interactions with the environment, s/he receives feedback on the actual state of the environment. A comparison is made of the feedback to the reference point. If the environment varies significantly from the reference point, the individual will assess his/her ability to change the environment to bring it into alignment with the reference point. This sequence of events iterates until either the environment is aligned with the reference point (through changes in the environment, the reference point, or both), or the individual deems actions to be futile -potentially resulting in negative cognitive, emotional, or behavioral consequences of the individual (Morris and Marshall 2004). -- INSERT [Figure 3.1] H E R E -Within RA-WAD's preference-elicitation method, users are repeatedly made to appraise the incongruence between the desired referent point (i.e., to have it all) and the actual state (i.e., tradeoffs must be made). Every attempt to avoid trading off an attribute by allotting a significant portion of importance weights to it brings the user negative feedback that his/her own act has taken him/her even further from the referent point, as s/he is now left with fewer importance weights to assign to other attributes. The repeated assignment of importance-weights (until the user reaches zero) creates the perception that s/he has a limited ability to reduce the gap by utilizing RA-WAD. As a result, the user will feel only limited control over manipulations of RA-WAD to achieve the goal. Conversely, R A - E B A users reduce the number of alternatives by attribute until they reach the final choice and thus feel that they are gradually aligning the current 34 environment with their referent point. This sequential practice engages users more effectively in the decision-making process, leading to an illusion of control defined as "a person's expectation of success on a task that is inappropriately higher than objective circumstances warrant" (Davis and Kottemann 1994). R A - E B A ' s guidance through this sequential step-by-step approach leads users to believe (albeit falsely) that they are approaching a solution to their problem, rather than to recognize that they are making a series of tradeoffs (Davis and Kottemann 1994; Payne et al. 1993).4 To summarize, the preference-elicitation method of RA-WAD, as compared to that of R A - E B A , does not provide users with adequate means to cope with attribute conflicts. As a result, RA-WAD users will feel less control than R A - E B A users do. Hla . RA-WAD users will perceive less control over manipulating R A to specify their preferences than R A - E B A users. 3.2.3 Perceived Quality of Recommendations Users will negatively relate the attribute conflicts experienced during the preference-specification phase to the perceived quality of recommendations. This is because an individual's experience of the decision-making process inevitably influences his/her perception of the quality of final outcomes (Dasgupta and Chanin 2000), which often fails to reflect the objective quality of such outcomes (Payne et al. 1993). Specifically, users' satisfaction with, and confidence in, final decision outcomes are 4 Some might argue that RA-EBA users may choose a high level of an attribute in an attempt not to trade off the attribute. Such a choice of the highest level eliminates the majority of alternatives, and hence they receive only a few recommendations. Therefore, even RA-EBA users cannot avoid confronting attribute conflicts. However, previous research has shown that EBA users frequently revise their choice of cutoff levels, presumably in order to find an appropriate level that leaves them with enough alternatives to choose from (Fasolo et al. 2005). In other words, they do not attempt to stay with the highest level once they realize it limits the number of recommendations. Rather, they make more conservative choices in order to leave more options (Swaminathan 2002). In addition, they are aware of merely choosing the least acceptable level, which means that the recommendations will include alternatives with more favourable levels than the chosen cutoff levels. 35 negatively influenced by any challenges, difficulties, and disagreements they experienced during previous decision-making processes (Aloysius et al. 2006; Kotteman and Davis 1991). Kotteman and Davis (1991) have argued that the decisional stress experienced during the preference-elicitation phase decreases users' confidence in the final outcomes proposed by the DSS. This claim parallels the findings in GDSS literature that disagreements experienced by group members during discussion sessions negatively affect their satisfaction with decision outcomes (Dasgupta and Chanin 2000). RA-WAD users feel that they have less ability to resolve attribute conflicts, because they are confronted with attribute conflicts during the preference-elicitation stage. In other words, RA-WAD users become uncertain about what the best course of action is, and about how to construct their preferences in response to attribute conflicts (Luce et al. 2001). Since their preferences may not have been clearly established in the previous stage, the question of which alternative options are better than others also becomes unclear. Therefore, RA-WAD users, who feel uncertain as to whether or not they have specified their preferences appropriately, are less likely to perceive that RA-WAD reflects their preferences correctly when generating recommendations. I expect that attribute conflicts highlighted by RA-WAD will lead to a low level of perceived quality of recommendations (see Figure 3.2). -- INSERT [Figure 3.2] H E R E -H2a. R A - W A D users will perceive the quality of recommendations less favourably than R A - E B A users. 3.2.4 Perceived Effort Attribute conflicts also increase consumers' perceived effort to use a given RA. A consumer facing attribute conflicts will choose to exert more effort in an attempt to solve 36 the problem (Schwarz, Higgins and Sorrentino 1986). Aloysius et al. (2006) argued that users evaluating a DSS that highlights conflicts tended to find it more demanding in terms of effort. In order to argue the negative effect of attribute conflicts on perceived effort, however, it is imperative that the two RAs require consumers of the same amount of objective effort to operate them. In other words, RA-WAD should not impose more effort on consumers to utilize it than R A - E B A . If so, consumers will perceive RA-WAD more effortful than R A - E B A regardless of an existence of attribute-conflicts. No effort was spared to equate the two RAs in terms of both computational operations and mental processes. M y N G O M S L analyses confirm that effort levels for RA-WAD and R A - E B A are equivalent (See Appendix 1). Given the equivalent amount of objective effort warranted, I posit that attribute conflicts will cause users to perceive RA-WAD more effortful than R A - E B A . H3a. R A - W A D users will perceive higher level of effort than R A - E B A users. 3.2.5 Usage Intentions Perceived control, quality of recommendations, and perceived effort as affected by RAs all influence decision-makers' intentions to (re)use RAs for future shopping (usage intentions, hereafter). Constructive decision making theory claims that consumers select the decision strategy that best supports the goals they want to achieve in a particular decision context, rather than invariantly applying the optimal strategy to any and all decision contexts (Bettman et al. 1998; Payne, Bettman and Johnson 1988). Consumers develop a hierarchy of goals that specifies which goals they want to achieve in a particular situation and then choose the approach they think will be most effective in attaining those goals. According to Bettman et al. (1998), the goals that capture many of 37 the most important motivational aspects are maximizing the accuracy of the decision, minimizing the cognitive effort required to make the decision, and minimizing negative emotions arising from attribute conflicts. RA-WAD and R A - E B A vary in their advantages and disadvantages with respect to accomplishing these goals in any given situation. Therefore, I argue that consumers will select the R A that best meets their goals for a particular situation, given the advantages/disadvantages of the RAs with respect to those particular goals. When their task does not involve attribute conflicts, consumers determining whether or not to use an R A consider only the levels of effort and quality that each R A involves. Decision-makers place higher value on effort-reduction than on accuracy-increase, because feedback on effort is more immediate than feedback on accuracy (Einhorn and Hogarth 1978; Kleinmuntz and Schkade 1993). The provision of a DSS reduces the effort associated with employing the compensatory strategy relative to other heuristics. Since the provision of a DSS satisfies the goal of effort-minimization, users switch to the compensatory strategy precisely because they can keep their effort level low, while also achieving high quality outcomes, as opposed to using the elimination heuristics (Todd and Benbasat 1991; Todd and Benbasat 1994; Todd and Benbasat 1996; Todd and Benbasat 1999). In brief, users place more value on effort-minimization than on accuracy-maximization. Accuracy-maximization begins to play a role only when effort is kept constant across decision strategies through the provision of DSS that reduces effort on the part of users. The presence of attribute conflicts modifies the dynamics of the effort-accuracy goal (Bettman et al. 1998; Luce et al. 1997). Consumers attempt to cope with the conflicts and this coping process shapes the goal hierarchy that specifies which goals they value more. Confronted with attribute conflicts, users place greater weight on the goal of minimizing such confrontations than on any other goals. Users prefer E B A that enables 38 them to avoid the knowledge that one attribute must be traded off for another to WAD that achieves higher accuracy but highlights the compromises that must be made (Luce 1998). In the current study, RA-WAD users, as compared to R A - E B A users, cannot avoid confronting attribute conflicts and perceive the level of effort higher and the quality of recommendations lower. Therefore, RA-WAD is inferior to R A - E B A in terms of accomplishing all three goals of maximizing accuracy, minimizing effort, and minimizing the experience of attribute conflicts. Therefore, H4a. R A - W A D users will show lower usage intentions than R A - E B A users. 3.2.6 Moderator Variable: Task Emotionality While attribute conflicts generally influence users' perceptions of, and intentions to use, RAs, this influence is moderated by the emotionality of the task being conducted. Task emotionality is defined as the degree to which a decision task is perceived to bring severe negative consequences to well-being of people who decision-makers care about (Luce et al. 1997; Luce et al. 1999). Task emotionality alters the user's primary goal in a particular decision context, and this goal inevitably influences user's choice of strategies (Bettman et al. 1993; Bettman et al. 1998), as follows. The prospect of tradeoffs arising from attribute conflicts becomes at once more unpleasant and more compelling for users facing high-emotion tasks. Consequently, for high-emotion tasks, a user's primary goal will be to minimize confronting attribute conflicts (Luce et al. 1997; Luce et al. 1999; Luce et al. 2000; Luce et al. 2001). RA-WAD does not help them to achieve this goal, whereas R A - E B A allows them to bypass the explicit confrontations with the prospect associated with the tradeoffs. Given the influence of task emotionality on users' goals, I claim that the negative effects of attribute conflicts on users' perceptions of RA-WAD will increase for a high-emotion task as compared to a low-emotion task, while they will 39 remain low with R A - E B A . In their empirical studies (1997, 1999), Luce et al. demonstrated the moderating effects of task emotionality on consumers' decision processes and strategy choices. According to Luce et al., "a decision task should be inherently emotional when severe negative consequences are possible, but the best course of action is unclear. " Luce et al. varied task emotionality to create high- and low-emotion groups, showing that task emotionality causes consumers to perceive their choice experience more negatively and leads them to choose E B A because it better enables them to cope with attributes conflicts (Luce et al. 1997; Luce et al. 1999). In their 1997 study, Luce et al. asked participants to select a child to receive financial support from a charitable organization. Participants in the high-emotion group were given snapshots of children who had received such support, as well as more specific and extensive background text describing the children's plight. Luce et al. believed that the more vivid information would enable decision-makers to fully apprehend the potential severity of the consequences of their upcoming decisions. Participants in the high-emotion group were then asked to imagine they had been supporting all five of the children in the decision set but now had to reduce their support to one child only by eliminating support for the other four children altogether. Participants in the low-emotion group were asked to imagine that they had previously supported no children and were now choosing one child who would begin to receive support. The results showed that the participants in the high-emotion group reported more negative emotions, and tended to choose the E B A strategy over the A C strategy more often than participants in the low-emotion group. Luce et al. (1999) extended their 1997 study by examining the moderating effects of task emotionality on consumers' typical purchase choices. To manipulate task 40 emotionality, they altered the reference points that described the decision task as either an advance or a deterioration from the current point, following Lazarus s (1991)'s notion that decisional stress results in part from appraisals regarding whether goals are likely to be furthered or blocked in a particular situation. Participants in the high-emotion group were told that they had previously rented an apartment in upper-average condition and were now moving down to one in only average condition, while participants in the low-emotion group were told that they had previously rented an apartment in lower-average condition and were now moving up to one in average condition. Participants in the high-emotion group viewed the decision task as a downgrade from their initial state, while those in the low-emotion group saw it as an improvement. Again, the results showed the moderating effects of task emotionality: participants in the high-emotion group experienced more decisional stress and therefore paid higher prices in order to avoid trading off an important attribute. Based on the research described above, I posit that task emotionality moderates the influence of attribute conflicts on users' perceptions of and intentions to use RAs for the following reasons. As task emotionality increases, attribute conflicts imply greater potential damage to the people consumers care about. Consumers executing a high-emotion task, therefore, experience more decisional stress around attribute conflicts than those executing a low-emotion task. RA-WAD not only fails to provide users with the perception that it supports their resolution of attribute conflicts but it also leaves them unsure about what actions to take in response to attribute conflicts. Since the failure to address attribute conflicts causes users more potential harm and losses in a high-emotion situation than in a low-emotion situation, consumers in a high-emotion situation will be less content with RA-WAD. Thus, they will evaluate RA-WAD more negatively as compared to R A - E B A . In contrast, R A - E B A leads users to believe that it helps them resolve attribute conflicts for both high- and low-emotion tasks. Therefore, users' 41 perceptions of R A - E B A will be less influenced by task emotionality. Consequently, the differences in users' perceptions and usage intentions between RA-WAD and R A - E B A are likely to be exacerbated for a high-emotion task but not for a low-emotion task. H l b . T h e d i f f e r e n c e i n p e r c e i v e d c o n t r o l b e t w e e n R A - E B A a n d R A - W A D u s e r s w i l l b e l a r g e r f o r a h i g h - e m o t i o n t a s k a s c o m p a r e d t o a l o w - e m o t i o n t a s k . H 2 b . T h e d i f f e r e n c e i n p e r c e i v e d q u a l i t y o f r e c o m m e n d a t i o n s b e t w e e n R A -E B A a n d R A - W A D u s e r s w i l l b e l a r g e r f o r a h i g h - e m o t i o n t a s k a s c o m p a r e d t o a l o w - e m o t i o n t a s k . H 3 b . T h e d i f f e r e n c e i n p e r c e i v e d e f f o r t b e t w e e n R A - E B A a n d R A - W A D u s e r s w i l l b e l a r g e r f o r a h i g h - e m o t i o n t a s k a s c o m p a r e d t o a l o w - e m o t i o n t a s k . H 4 b . T h e d i f f e r e n c e i n u s a g e i n t e n t i o n s b e t w e e n R A - E B A a n d R A - W A D u s e r s w i l l i n c r e a s e f o r a h i g h - e m o t i o n t a s k b u t n o t f o r a l o w - e m o t i o n t a s k . 3.3 RESEARCH METHOD I conducted a laboratory experiment in order to test empirically the main effect of R A types on consumers' perceived control, quality of recommendations, effort, and usage intentions, as well as the interaction effect between R A type and task emotionality. The experiment allowed the exercise of control over independent, moderating, dependent, and possibly confounding variables to achieve a high degree of internal validity (Singleton and Straits 1999). A 2 x 2 factorial design with two between-subject factors was used. The first factor, R A types, has two levels: RA-WAD and R A - E B A ; and the second factor, task emotionality, also had two levels: high-emotion and low-emotion. 3 .3 .1 R A D e s i g n RA-WAD and R A - E B A were developed with identical Web interfaces except for 42 their preference-elicitation methods. Each introduction page contained a welcoming message and a brief introduction to the R A consumers were about to use. The R A then presented participants with the list of attributes. Upon clicking an attribute, participants were prompted (1) to distribute 100 constant-sum importance weights (in the case of RA-WAD, see Figure 3.3), and (2) to specify the cutoff level (in the case of R A - E B A , see Figure 3.4). After specifying their preferences for attributes they considered important in their purchases, participants clicked the "View Results" button, and this brought up a list of recommendations. On this page, participants could click a "Modify My Answers" button to return to the preference-elicitation pages if they were not satisfied with the recommendations provided. - INSERT [Figure 3.3 and 3.4] H E R E -3.3.2 Alternatives and Attributes I chose used cars as the product type for this study since the attributes of used cars are closely related to the well-being of passengers, such that attribute conflicts signal potentially severe losses to decision-makers. Used cars have been employed by previous studies in which task emotionality was investigated (Luce 1998; Luce et al. 1997). In order to simplify the decision task, I chose only four-door sedans with automatic transmissions, as the inclusion of other vehicle types (e.g., SUVs, trucks, or vans) would have prohibited one from describing products with the same set of attributes. Table 3.1 shows the attributes of used cars. Used cars have nine attributes: breakdown rate of engine, brakes, transmission, cooling systems, exhaust systems, fuel systems, crash test results, safety features, and price. Each attribute has five levels. These attributes and levels were drawn from well-known consumer reviews, such as Consumer Reports (www.consumerreports.org) and the Kelley blue book (www.kbb.com). The main purpose of choosing these attributes was to induce participants to perceive the sense of 43 dangers and harms that could occur to the people they cared about. In other words, even if participants were not particularly knowledgeable about each of these attributes, they could easily expect potential negative results from the "breakdown" of each part. In addition, every effort was made to ensure that participants were made aware of these consequences: During the experimental sessions, participants were required to familiarize themselves with these attributes. The RAs provided a link that opened a pop-up window containing detailed description about each attribute (See the link, "View details of [an attribute], in Figures 3.3 and 3.4). Each description warned the participants of potential consequences that could occur if the particular attribute were to malfunction. For instance, the description about breaks breakdown rate included, "If the brakes fail, stopping your car will be delayed or fail, and hence the result can be disastrous." - INSERT [Table 3.1] H E R E -The condition of attribute conflicts was achieved by altering the degree to which the favourable values of attributes were related to the unfavourable values of other attributes. A l l attributes were negatively correlated with price. For instance, the attainment of a higher (more desirable) level of engine reliability resulted in a $500 increase in price. I ensured that such inter-attribute conflicts were noted by participants by providing decisional guidance describing inter-attribute conflicts (see Figures 3.3 and 3.4). Decisional guidance has previously been used to inform consumers of the potential costs of obtaining advanced product features (Wang and Benbasat 2007). In this study, decisional guidance was used to explain specific vehicle attributes and the conflicts between these attributes and price. 3.3.3 Manipulation of Task Emotionality Task emotionality can be altered by varying either (1) the decision problem itself, 44 i.e. products or attributes and attribute conflicts, or (2) the decision task (Luce et al. 1997; Luce et al. 1999). Varying decision problems across task emotionality levels confound task emotionality with other cognitive factors. First, altering products or attributes makes it extremely difficult to keep product characteristics and consumers' attitudes towards products constant across high/low-emotion groups. Two distinct products are unlikely to have exactly the same number of attributes of equal importance to consumers (e.g., automobiles vis-a-vis MP3 players). In addition, consumers' attitudes to two distinct products - their level of involvement with each product, their sense of its particular relevance to them - inevitably differ. Consequently, altering product attributes confounds task emotionality with product/consumer characteristics. Similarly, Widing and Talarzyk (1993) have pointed out that varying attribute conflicts also alters the attractiveness of the alternative sets. When attributes are positively correlated, more alternatives are likely to seem favourable to consumers (Widing and Talarzyk 1993). The apparent favourableness of alternative sets influences users' perceptions of RAs. As a result, the characteristics of alternative sets also are confounded with task emotionality (Widing and Talarzyk 1993). In order to avoid these issues, Luce et al. (1997, 1999) altered decision tasks instead of altering decision problems per se in their studies. Firstly, they increased the vividness of the described negative decision consequences by providing more cues illustrating potential consequences. Secondly, they altered the reference points that framed the upcoming choices as either improvements or downgrades from the products currently owned. Following Luce et al. (1997, 1998), I manipulated task emotionality by altering (1) the vividness of the described negative decision consequences and (2) the reference points. 45 First, participants in the high-emotion task group were asked to watch a video clip of interviews with a family who had lost their young son in a car crash. This video directly alerted participants to the potentially tragic consequences of their upcoming decisions. Meanwhile, participants in the low-emotion task group watched a different video clip that conveyed general, non-emotional messages, such as tips for safe driving for a summer vacation. Next, all participants were provided with either a high or low reference point for their decisions (see Appendices A.2 and A.3). Participants were told that they only had a budget high enough to buy a used car in average condition. Those in the high-emotion group were asked to imagine that the car which their family currently drove was in higher-than-average condition, whereas those in the low-emotion group were asked to imagine that it was in lower-fhan-average condition. Thus, participants in the high-emotion group were expected to find it more difficult to trade off attributes, given that these tradeoffs would result in a downgrade from what they already owned (i.e., they must decide how much of the attributes they could give up); conversely, participants in the low-emotion group were expected to find it easier to trade off attributes because they were deciding how much of the attributes they would improve. In order to check whether or not the manipulation of task emotionality was successful, I refined Drolet and Luce (2006)'s 7-point items, which assesses the severity of the potential consequences of the upcoming decision and the degree of threat that is associated with the decision task. I included two additional scale measures in order to evaluate the cognitive difficulties of the decision tasks: (1) how important they expected their decision to be and (2) how challenging they expected their decision to be (Luce et al. 1997). My goal was to alter the emotional aspect of the task while also keeping the cognitive aspect (such as, importance and cognitive difficulty) to be constant, in order to single out attribute conflicts as the sole moderator on the dependent variables. I therefore 46 expected these additional measures to show no difference between the high- and low-emotion groups. 3.3.4 Operationalization of the Dependent Variables A l l four dependent variables were measured on 7-point scales (see Table 3.2). Bechwati and Xia (2003) 's three control scales were revised: control felt during specifying preferences for used cars (two items), and whether the RA's preference-elicitation method made users feel in control. Widing and Talarzyk's (1993) perceived accuracy measure was modified with decision quality replaced to recommendation quality. Perceived quality of recommendations included four items: whether the best used cars are included in the recommendation set, whether the used cars recommended matched their needs (normal and reversed scale), and whether they would choose from the same set of alternatives on a future occasion. Perceived effort is conceptualized based upon Bechwati and Xia (2003)'s four effort scales: the degree to which use of R A required time and effort, and the ease and complexity of RAuse. Lastly, intention-to-use a system in the future was conceptualized based upon the measure proposed by Pavlou (2003) and Venkatesh and Davis (2000), which gauges a respondent's intention-to-use a system when they are given access to the system in the future. - INSERT [Table 3.2] H E R E -3.3.5 Participants and Experimental Procedures A total of 100 students at a large North American university participated in the experiment. 25 participants were randomly assigned to each of the four treatment groups. The participants' previous knowledge of the product category was carefully controlled by recruiting only those who had (1) a driver's license, (2) no experience of purchasing a used car, and (3) moderate expertise with automobiles (only those who indicated 2-5 on a 7-point automobile expertise scale ranging from 1 [not at all expert] to 7 [extremely Al expert]). It was important to recruit participants with only moderate knowledge of used cars because only those without sufficient knowledge of the product, and without clearly established preferences towards the product, would seek advice from RAs (Haubl and Trifts 2000; Swaminathan 2002). Furthermore, as Swaminathan (2002) has pointed out, consumers with greater knowledge are better able to make trade-offs between attribute levels than consumers with less knowledge. Therefore, people who indicated 6 or 7 on the automobile expertise rating were not recruited. However, it was desired that participants were not complete novices with regard to automobiles because they would then have lacked the basic understanding of the product that would enable them to grasp the negative consequences associated with automobile malfunctions and traffic accidents (Swaminathan 2002). Therefore, participants who did not have a driver's license (hence lacked any previous driving experience), and who indicated 1 on a 7-point automobile expertise scale, were excluded. Table 3.3 shows the demographic factors of the participants. - INSERT [Table 3.3] H E R E -The experiment session proceeded as follows. Participants watched a video clip about automobiles (i.e., a clip containing a car crash in the high-emotion group and a clip containing driving tips for summer vacations in the low-emotion group) and were then given a written decision task that required them to select a used car for one of their family members. Prior to reading the task, they were told to think of one of their family members. They were asked to imagine as hard as they could that the chosen family member was facing the situation described in the task and that they were selecting a used car for this person. The rationale behind specifically involving a family member is that consumers feel more stressed when trading off important goals that may affect the well-being of someone they care about (Tetlock et al. 2000). In addition, participants were made to watch the video clip prior to reading the task, so that they would consider the 48 potential consequences of associated with automobile malfunctions and traffic accidents portrayed in the video clip while reading the task. Once participants had read the task, they were asked to answer the four questions for the manipulation check online. Next, each of the participants was trained in how to use an R A and how to navigate the Web interfaces. To ensure identical training, the same training video was used for all participants. Once participants fully understood how to. manipulate the R A and the interfaces, they were asked to use the R A freely to choose a used car. They were allowed to spend as long as they wanted doing this, and to modify their preferences as many times as they liked. They were also allowed to refer to the task sheet while making their choices. Once participants had made their final decision, they were asked to complete an online questionnaire containing the four dependent variables. Each participant was guaranteed monetary compensation for his/her participation ($20). In order to motivate participants to consider the experiment as a serious online shopping session and to increase their involvement, the top 25% performers were offered an additional incentive of $40. Participants were told before the experiment that they would be required to justify their choices and that their performance would be judged based on these justifications. The main criterion for judgment was the extent to which a participant's justifications appropriately and convincingly supported his/her choice of the used car selected. 3.4 R E S U L T S A N D A N A L Y S E S SPSS for Windows Version 14.0 was used to conduct analyses of all the variables for both manipulation checks and dependent variables. 3.4.1 Manipulation Checks Participants in the high-emotion group believed that their upcoming decisions 49 potentially carried more severe consequences (mean = 5.66) and were potentially more threatening (mean = 4.86), compared to participants in the low emotion group (mean = 4.98 for severity; mean = 4.12 for threats). The difference in the composite measure of task emotionality (aggregating the two items) between the two emotion groups was statistically significant (t = 2.87,p < .05). In contrast, participants in both emotion groups assessed the tasks as equally important (mean = 6.62 for the high-emotion group; mean = 6.46 for the low-emotion group) and challenging (mean = 6.12 for the high-emotion group; mean = 6.10 for the low-emotion group). The difference in the cognitive aspect of the tasks (aggregating the two items) was not significant (t = .910, p > .05). Clearly, the two tasks varied in terms of emotionality, but not in terms of cognitive importance and difficulty, showing that task emotionality was manipulated successfully and independently from cognitive aspect of the tasks. 3.4.2 Testing of Hypotheses Validity of all the dependent variables were tested and deemed appropriate by previous studies (See the sources of the variables in Table 3.2). In this study, Cronbach's Alpha was used to assess the reliability of the measures, control, recommendation quality, effort, and usage intentions. A l l constructs displayed acceptable levels of Cronbach's Alpha, .85, .83, .79, and .95, respectively (see Table 3.4). -- INSERT [Table 3.4] H E R E -In order to prevent any extraneous factors from influencing the results, I measured and controlled variables known to influence users' adoption of RAs. These control variables entail consumers' current web usage frequency, previous web experience (Taylor and Todd 1995), gender (Gefen and Straub 1997), tendency to pursue accuracy over effort (Wang and Benbasat 2007), and involvement with used cars (Bechwati and Xia 2003). 50 Table 3.5 shows the means and standard deviations of the dependent variables. The results are depicted in Figures 3.5, 3.6, 3.7, and 3.8. - INSERT [Table 3.5 and Figure 3.5,3.6, 3.7, and 3.8] H E R E -A N C O V A was run to analyze the main effect of R A type, and the interaction effect between R A type and task emotionality on each dependent variable (Refer to Tables 3.5, 3.6, 3.7 and 3.8). The RA type main effect was significant on perceived control (F(l , 91) = 9.469, p <.01), recommendation quality (F(l , 91) = 35.346, p <.01), and usage intentions (F(l , 91) = 7.105, p <.01). These results support hypotheses H l a , H2a, and H4a, which predicted the main effect of RA-WAD on perceived control, recommendation quality, and intentions-to-use. However, the main effect of RA type was not significant on perceived effort (F(l , 91) = 1.420, p >.05). Thus, H3a, which anticipated the main effect of RA type on perceived effort, is not supported. - INSERT [Table 3.6, 3.7, 3.8, and 3.9] H E R E -The R A type x task emotionality interaction effect was not significant on perceived control (F(l , 91) = .978, p >.05). Therefore, H lb , which predicted that the moderating effect of task emotionality on perceived control, is not supported. However, the difference in the mean scores of perceived control between R A - E B A and RA-WAD was greater (0.82) for the high-emotion task than that for the low-emotion task (0.43) (see Figure 3.5). This result shows a possibility that the moderating effect on perceived control is existent. Therefore, I conducted two separate t-tests comparing the mean difference between RA-WAD and R A - E B A for each high- and low-emotion task. I applied a Bonferroni family adjustment to the level of significance in order to control for an inflated Type I error. The level of significance applied to the separate post-hoc t-tests is .05/2 or .025. In the high-emotion condition, the difference in perceived control between RA-WAD and 51 R A - E B A was significant (t = -3.064, p < .025). In the low-emotion condition, the differences was not significant (t = -1.591, p > .025). The t-test results indicates that the R A type x task emotionality interaction effect on perceived control is existent. In other words, RA-WAD reduced users' perceived control only in the high-emotion condition but not in the low-emotion condition. The R A type x task emotionality interaction effect was significant on recommendation quality (F(l , 91) = 6.675, p < .05) and usage intentions (F(l , 91) = 5.757, p < .05). Therefore, H2b and H4b, which anticipated the R A type x task emotionality interaction effect on recommendation quality and usage intentions, are supported. However, the R A type x task emotionality interaction effect was not significant on perceived effort, rejecting H3b (F(l, 91) = .076, p > .05). Like perceived control, I conducted two separate t-tests comparing the mean difference in perceived effort between RA-WAD and R A - E B A for each high- and low-emotion task. Again, a Bonferroni family adjustment was applied to the level of significance and the level of significance applied is .05/2 or .025. In neither of the high- and low-emotion condition, the values of perceived effort were significantly different between the RA-WAD and RA-E B A groups (t = -.792, p > .025 [high-emotion condition]; t = .608, p > .025 [low emotion condition]). In summary, the results indicate that RA-WAD, as compared to R A - E B A , decreases users' perceived control, recommendation quality, and usage intentions. The decrease in recommendation quality and usage intentions is more pronounced for high-emotion tasks than for low-emotion tasks. Although RA-WAD was generally perceived to provide lower perceived control than R A - E B A regardless of task emotionality, RA-WAD reduces perceived control more in the high-emotion condition than in the low-emotion condition. RA-WAD and R A - E B A require the same amount of effort regardless of task emotionality. 52 3.4.3 Relationships among the Dependent Variables The A N C O V A results show that attribute conflicts and task emotionality affect perceived control, recommendation quality, and usage intentions as hypothesized, but not perceived effort. The question remains, however, as to how perceived control, recommendation quality, and effort influence usage intentions. To answer this question, I used structural equation modeling techniques. According to the Constructive Decision Making Theory by Bettman et al. (1998), when attribute conflicts do not exist, users' primary motivations for choosing particular decision strategies are: (1) to maximize accuracy of decision and (2) to minimize the effort to make decisions. The emphasis on effort-accuracy changes as attribute conflicts come into a picture (Luce 1998; Luce et al. 1997; Luce et al. 1999; Luce et al. 2000). Consumers' primary motivation then becomes to reduce confronting attribute conflicts at the expense of effort. Consumers seek to spend more effort to justify to themselves and others that they are trying their best not to trade off important attributes (Luce et al. 2001). In this study in which attribute conflicts are primary elements of the decision tasks, perceived effort should not influence usage intentions, as consumers now focus more on a way to deal with attribute conflicts than to reduce effort. This assertion is in accordance with Technology-Acceptance-Model (TAM) literature claiming that perceived ease-of-use is not an important predictor of usage intentions when consumers use a system to achieve a clear utilitarian goal; perceived ease-of-use is influential when consumers want to achieve a hedonic goal, such as navigating websites without a clear goal to achieve (Koufaris 2002). Take together; I expect that perceived effort is not related to usage intentions. Quality of recommendations is positively related to usage intentions. It is important to note again that quality of recommendations refers to consumers' subjective perceptions of the recommendations, which often do not reflect objective accuracy of 5 3 recommendations (Todd and Benbasat 1999). Consumers dealing with attribute conflicts tend to choose the E B A strategy in part because they overestimate the accuracy of outcomes that result from E B A use (Luce et al. 2001). Consumers tend to be confident in the choices that they can rationalize with objective facts (Luce et al. 2001). E B A enables consumers to justify their choices on the basis of objective facts (e.g., "I had a budget of $5000; so I chose this particular price level; then the R A gave me only the vehicles that did not have comprehensive safety features. So, I had to choose one that had fewer safety features.). Such overestimated belief in accuracy of decisions leads them to choices of E B A (Luce et al. 2001). As a consequence, it is expected that consumers' (subjective perceptions) of recommendation quality positively influences their intentions to use the RA. Perceived control is positively related to recommendation quality and negatively related to perceived effort. As argued earlier, challenges, difficulties, and disagreements experienced during decision making processes negatively influence consumers' satisfaction with and confidence in final decision outcomes (Dasgupta and Chanin 2000; Davis and Kottemann 1994). RA-WAD's preference-elicitation method manifests attribute conflicts to the users without providing a clear way to resolve the conflicts; hence, the users would feel less satisfied with the recommendations provided by RA-WAD. Similarly, users would feel a R A more complex, difficult, and effortful when they did not feel that the R A was under their control: i.e., they could not manipulate the R A in their discretion due to the limited importance weights. In short, sense of control induces users to feel that the system is easy to use (Averill 1973; Frese et al. 1987). Perceived control indirectly influences usage intentions through recommendation quality, rather than influencing it directly. Previous studies, such as Shneiderman (1997), DeLone and McLean (2003), and Dasgupta and Chanin (2000), claimed indirect effect of perceived control on usage intentions through task performance and confidence in 54 decision outcomes, the constructs closely related to recommendation quality, rather than direct effect on usage intentions. That is, these arguments posit that perceived control affects consumers' perceptions of the quality of recommendations, which in turn influences their usage intentions. Based upon these previous studies, I argue that recommendation quality mediates the effect of perceived control on usage intentions. 3.4.4 Structural Equation Model Analysis Structural equation modeling procedures, implemented in PLS Graph 3.0, were used to perform a simultaneous evaluation of both the quality of measurement (the measurement model) and construct interrelationships (the structural model). PLS Graph provides the ability to model latent constructs even under conditions of non-normality and small- to medium-sized samples (Chin, Newsted and Hoyle 1999). In addition, the use of structural equation modeling techniques for a mediation analysis has major advantages as compared to other methods, such as regressions: It directly tests all paths and incorporates complications of measurement error, correlating measurement error directly into the model (Baron and Kenny 1986). The sample of 100 cases is adequate for PLS analysis. It satisfies the requirement that the sample size be at least 10 times the largest number of structural paths directed at any one construct. The largest number of paths to any construct in the research model is three: the three paths from perceived control, effort, and recommendation quality to usage intentions (see Figure 3.9). - INSERT [Figure 3.9] H E R E --Internal consistency was assessed by examining Cronbach's Alpha for each construct and the Composite Reliability produced by PLS (Table 3.4). Each of the measurement items loaded on its latent constructs significantly at .001 level, indicating high individual item reliability (Table 3.4). Both Cronbach's Alpha and the composite 55 reliability for all constructs were above the suggested threshold of .7, which is considered the benchmark for acceptable reliability (Barclay, Thompson and Higgins 1995). Gefen and Straub (2005) have suggested two criteria to examine discriminant validity. The first criterion requires the correlation of the latent variable scores with the measurement items' needs to show an appropriate pattern of loadings, one in which the measurement items load highly on their theoretically assigned factor and not highly on other factors. Specifically, all the loadings of the measurement items on their assigned latent variables should be an order of magnitude larger than any other loading (Gefen and Straub 1997). The factor- and cross-loadings reported in Table 3.10 demonstrate adequate discriminant validity. Each item loads highly on its latent variable and less on other variables. The second criterion requires that the square root of every A V E be much larger than any correlation among any pair of latent constructs and should be at least .50. Table 3.4 shows that this criterion was satisfied by the current data. - INSERT [Table 3.10] H E R E -The first diagram in Figure 3.9 shows the results of path analyses. As expected, the path coefficient between recommendation quality and usage intentions is positive and significant (t = 2.475, p < .05) indicating that recommendation quality has a positive impact on usage intentions. However, the path from perceived effort to usage intentions is negative but not significant (t = 1.544, p > .05) confirming the expectation. Perceived control positively influences recommendation quality (t = 9.697, p < .01) and negatively influences perceived effort (t = 5.179, p < .01), but does not directly influence usage intentions (t = 1.484,/? > .05). This result provides a partial support that perceived control indirectly influences usage intentions through recommendation quality. To further validate the indirect effect of perceived control on usage intentions, I followed Hoyle and Kenny's (1999) two procedures, commonly used to establish mediation effect. 56 Hoyle and Kenny (1999) claim that mediation can be profitably represented using a pair of path diagrams. Statistical evidence of mediation requires the following: (1) evidence of a causal influence of the Independent Variable (IV) on the Dependent Variable (DV), reflected statistically as a nonzero value for the path coefficient between IV and DV; and (2) a significant indirect effect of IV on DV, reflected statistically as a nonzero value for the product of the coefficient between IV and the mediator and between the mediator and DV, indicative of a decline in the direct effect of IV on D V when the mediator is accounted for (Hoyle, Kenny and Hoyle 1999). Sobel's test is often used to verify the significance of the indirect effect (Sobel and Leinhart 1982). If the indirect effect is significant and the direct effect is insignificant when the mediator is taken into account, then the mediator fully mediates the effect of IV on DV. The second diagram in Figure 3.9 shows that perceived control influences usage intentions, t = 3.445, p < .01, when the effect of perceived effort on usage intentions is controlled. Hence, the first condition for the mediation effects is satisfied. When recommendation quality is entered, as the first diagram shows, both the path from perceived control to recommendation quality (t = 12.103, p < .01) and the path from recommendation quality to usage intentions (t = 2.756, p < .01) are significant. In addition, as shown in the first diagram, the direct effect of perceived control on usage intentions becomes insignificant when recommendation quality is accounted for. The coefficient of perceived control is insignificant after recommendation quality is added to the diagram (t = 1.253, p >.05). Sobel's (1982) test results confirmed that the indirect effect was significant (t = 2.687,/? < .01), satisfying the second condition. Given that the two steps were satisfied, I conclude that the effects of perceived control on usage intentions are fully mediated by recommendation quality. The results of the mediation analyses indicate that when users do not perceive that RAs provide them with an approach (i.e., preference-elicitation method) with which they 57 can resolve attribute conflicts at their own discretion, they are less satisfied with the recommendations provided by the RA, which in turn lowers their intentions to use the RA. 3.5 DISCUSSION A N D C O N C L U S I O N S 3.5.1 Summary of Findings Purchase decisions are particularly stressful when product attributes are negatively correlated, such that consumers must trade off one attribute in order to attain another. RA-WAD, as compared to R A - E B A , does not support users' coping with emotion-laden attribute conflicts, thereby negatively influencing decision-makers' R A perceptions and usage intentions. In addition, task emotionality alters the extent to which users feel the need to cope with attribute conflicts. Users are more motivated to cope with high-emotion tasks, which are associated with severe negative consequences, than low-emotion tasks, which they connect with little harm to themselves. As a result, the negative effects of RA-WAD on users' perceptions and usage intentions are more pronounced for high-emotion tasks, while these effects are limited in regards to low-emotion tasks. R A - E B A , meanwhile, enables users to avoid confrontation with attribute conflicts for both high-and low-emotion tasks. Therefore, the differences between RA-WAD and R A - E B A are greater for a high-emotion task than for a low-emotion task. These hypotheses are supported by the experiment except the hypotheses regarding perceived effort. The results of a structural equation model indicate that perceived control influences usage intentions indirectly through recommendation quality. 3.5.2 Discussion of the Results In this section, I discuss the results obtained from A N C O V A , in conjunction with the structural equation model; then present possible explanations for the hypotheses not supported. 58 The A N C O V A results show that use of RA-WAD decreases consumers' perceived control, recommendation quality, and intentions to use the RA, as compared to R A - E B A . The structural equation model explains how perceived control exerts influence on recommendation quality, thereby affecting usage intentions. More specifically, RA-WAD decreases consumers' perceived control, as compared to R A - E B A (as shown in the A N C O V A results). Such decrease in perceived control leads to a decrease in perceived quality of recommendations, which in turn results in lower usage intentions (as shown in Figure 3.9, the structural equation model). In other words, consumers become less intent on using RA-WAD than R A - E B A , because use of RA-WAD, as compared to use of RA-E B A , does not provide them with the sense of control (in using the R A to resolve the attribute conflicts) with which, they believe, they can attain higher quality recommendations. Most of the hypotheses for this study were supported, with two exceptions: (1) the non-significant main effect of attribute conflicts on perceived effort and (2) the non-significant interaction effect between R A type and task emotionality on perceived effort. These non-significant results meant that neither attribute conflicts nor task emotionality increased users' perceived effort in this experiment, contradictory to Schwarz (1986)'s notion that individuals facing attribute conflicts choose to exert more effort to solve the given problem (Schwarz et al. 1986). It was expected that RA-WAD users would perceive higher level of effort than RA-E B A users. Indeed, A N C O V A results, conducted with the same control variables noted earlier, show that RA-WAD users (mean = 7.84, SD = 7.14) revised their preferences more often than R A - E B A users (mean = 4.28, SD - 2.77), F(l,91)=11.633, p < .01, as we expected they would in light of attribute conflicts. Given that RA-WAD does not require more computational and mental effort to use than R A - E B A assuming that they set their attribute levels once before they view their recommendations (see N G O M S L results in 59 Appendix 1), the increased number of iterations in revising attribute levels by RA-WAD would have led to higher objective effort. If this is the case, why was this difference not reflected in perceptions of effort spent? Clearly, the increase in objective amount of effort was not reflected in users' subjective perception of effort expenditure. The non-significant results may be attributed to the incongruence between perceived effort that users report and objective effort that users actually have exerted. The reason for the incongruence may lie in the fact that users spent more effort not because RA-WAD mandated so, but because they tried to resolve attribute conflicts at their own discretion. There is strong evidence in the literature to support this contention. System-users' perceived effort level does not parallel the objective effort they actually spend especially when they exerted more effort at their own will (Agarwal and Karahanna 2000). Moreover, individuals' voluntary effort to improve outcome quality is viewed differently from the compulsory effort to use a system (Pereira 2001). In other words, users can justify to themselves the reason they exerted effort to use a system - whether because the system is so complicated that demands users' effort, or it is because it is themselves who deliberately exert effort in order to improve the quality of outcomes that will accrue by the system use. We believe that the latter is the case, that is, especially in the face of attribute conflicts the users decided to invest more in making a more accurate decision by going through multiple iterations; hence, this explains the non-significant difference in their perceptions of effort. 3.5.3 Limitations This study has a number of limitations. First, investigating task emotionality in a laboratory has clear limitations. However, given the importance of emotionality in this study's context, it was necessary to conduct the experiments in highly controlled laboratory settings. I was able to manipulate levels of task emotionality and achieved the predicted interaction effects between task emotionality and R A types. However, I do not 60 maintain that our obtained levels of task emotionality can imitate the intensity of high-stakes decisions in the real world. Second, manipulation of task emotionality was accomplished by combining the provision of vivid descriptions of negative consequences and altering reference points. This manipulation, based on previous studies in which researchers investigated the effect of task emotionality, made it impossible to separate the effects of each element. However, the purpose of this study was not to discover which factor causes task emotionality, but rather to examine the influence of task emotionality on users' perceptions of RAs. 3.5.4 Contributions and Implications This study has important contributions to theory advancement over and beyond the prior studies on attribute conflicts, such as Aloysius et al. (2006). Firstly, I have extended R A literature from the effort-accuracy perspective to the area of attribute conflicts, and shown that the RA-WAD that provides accurate results is not necessarily preferable to R A - E B A , unless the problem of attribute conflicts is resolved. The result suggests the importance of investigating attribute in the studies of RAs for products with high emotional content. Secondly, this is the first study in which the negative effects of attribute conflicts on perceived control have been examined. The inclusion of perceived control enhances our understanding of why product-attribute conflicts negatively influence consumers' perceptions of the RA. Thirdly, the study shows that investigations into consumers' perceptions of RAs should consider the emotionality of the task supported by the RAs, as task emotionality determines the extent to which users value the RA's support in coping with attribute conflicts. Clearly, consumers' perceptions of RAs do not remain invariant but change contingent upon the nature of the tasks being conducted. Lastly, I overcome a limitation in previous studies, in which attribute conflicts were manipulated together with the cognitive effort required to perform the tasks. In the current study, I managed to alter task emotionality while keeping the cognitive 61 complexity of tasks (i.e., number of attributes and alternatives) constant across the emotion groups. This study also has many implications to practice. To date, R A designers have strived to. develop the algorithms of RAs that provide highly accurate recommendations. However, Study 1 demonstrated that consumers' perceptions of recommendations do not reflect the objective accuracy of recommendations for emotion-laden products. Such an overestimation of accuracy resulted from using the R A - E B A has been found consistently in many empirical studies, including Kotteman et al. (1994), Aloysius et al. (2006), and Todd and Benbasat (1999), to name a few. This consistent result suggests that a further refinement of RA's algorithms may not change the consumer's fundamental biases (Griffin and Tversky 2002). Furthermore, i f the consumer is satisfied with the R A and the decisions made by using the RA, then, probably that is what it counts. A substantial portion of consumers pursue not only making the most accurate decisions but also enjoying their shopping experience (Armstrong and Kotler 2004). This suggests that the designer's primary goal should be not just to develop the RA ' s algorithms that increase objective accuracy but to devise the R A that is perceived to be accurate. Therefore, R A -E B A may be a good option for products with high emotional content. 3.5.5 Future Research In a future study, researchers may investigate how the preference-elicitation method of RA-WAD should be designed to alleviate its conflict-confronting nature, so as to encourage consumers to use this desirable strategy. Researchers may also want to take a closer look at perceived control, given that the results of this study suggest that it is not actual control but an illusion of control that determines users' confidence in dealing with attribute conflicts. In fact, R A - E B A can be viewed as more restrictive, considering that not only do 62 attribute conflicts prevent a user from choosing the highest level of an attribute (which results in the elimination of the majority of alternatives), but also the emotional nature of the task stops the user from choosing the lowest level (because choice of the lowest level indicates that the user is trading off the attribute closely related to the well-being of the people they care.). Therefore, users are left with only a few preference options per attribute to eliminate alternatives. For instance, i f the attribute has five levels, the levels the user can actually choose are only 2-4, excluding the highest and the lowest levels. In contrast, users can calibrate RA-WAD in many different ways by allotting many different combinations of importance weights. As such, the relationship between restrictiveness and perceived control within a choice situation involving attribute conflicts may be an interesting topic for a future study. 63 3.6 T A B L E S A N D F I G U R E S Table 3.1 Attributes of Used Cars Attributes Levels Breakdown rate of engine Breakdown rate of brakes Breakdown rate of transmission Breakdown rate of cooling systems Breakdown rate of fuel systems Breakdown rate of exhaust systems (1) Excellent (5.0% or fewer chances of developing problems within next few years) (2) Good (5.0% to 20.0% chances of developing problems within next few years) (3) Average (20.0% to 35% chances of developing problems within next few years) (4) Mediocre (35% to 50% chances of developing problems within next few years) (5) Poor (more than 50% chances of developing problems within next few years) Crash test results (1) Excellent (probably no injury or a minor injury) (2) Good (moderate injury likely) (3) Average (certain injury, possibly severe) (4) Mediocre (severe or fatal injury highly likely) (5) Poor (severe or fatal injury virtually certain) Safety features (1) Excellent (equipped with antilock brakes, airbags [front, side, and head-protection] for all passengers, and safety belts5) (2) Good (equipped with air bags [front, side, and head-protection] for the driver and the right front passenger and safety belts) (3) Average (equipped with air bags [frontal protection only] for the driver and the right front passenger and safety belts) (4) Mediocre (equipped with safety belts only) (5) Poor (equipped with damaged or malfunctioning safety belts) Price (1) $875-$2,000 (2) $3,000 - $4,500 (3) 4,500 - $6,000 (4) $6,000 - $7,500 (5) $7,500-$19,300 5) Safety belts are for all passengers. 64 Table 3.2 Dependent Variables Variables Measures Sources Perceived Control When specifying my preferences for used cars, 1 felt 1 was in control. 1 think that 1 had a lot of control over the preference-specification process. The way 1 indicated my preferences for used cars made me feel 1 was in control. Bechwati and Xia (2003) Recommendation Quality Used cars that suit my preferences were recommended by the virtual advisor. Used cars that best match my needs were provided by the virtual advisor. The used cars recommended by the virtual advisor did NOT match my needs ®. 1 would choose from the same set of alternatives provided by the virtual advisor on my future purchase occasion. Widing and Talarzyk (1993) Perceived Effort The task of using the virtual advisor to choose a used car took too much time. Using the virtual advisor to choose a used car required too much effort. The task of using the virtual advisor to select a used car was easy ®. The task of using the virtual advisor to select a used car was too complex. Bechwati and Xia (2003) Usage Intentions Assuming 1 have access to the virtual advisor, 1 intend to use it next time 1 consider buying a used car. Assuming 1 have access to the virtual advisor, 1 predict 1 would use it next time 1 plan to purchase a used car. Assuming 1 have access to the virtual advisor, 1 plan to use it next time 1 consider buying a used car. Wang and Benbasat (2004) 65 Table 3.3 Demographics of Participants N Minimum Maximum Mean Std. Deviation Age 100 18.00 33.00 20.9600 2.386 Number of years using the Web 100 4.00 15.00 8.6300 2.082 Product Involvement 100 1.00 7.00 3.9975 1.527 Current Web Use Frequency* 100 1.00 6.00 2.6100 .956 Tendency to pursue accuracy over effort 100 3.00 7.00 5.3167 .946 * The five options for the current web use frequency were: (1) less than 30 minutes/week, (2) 1-2 hours/week, (3) 2-4 hours/week, (4) 4-8 hours/week, and (5) more than 8 hours/week Table 3.4 Means, S D s , Inter-Construct Correlations and Average Variance Extracted Mean (SD) Cron bach's Alpha Composite Reliability (D (2) (3) (4) Effort 2.315 (.846) .797 .871 .792 a Control 5.180 (.994) .859 .916 -.463°** .885 Recomm. Quality 5.230 (.946) .836 .896 -.430** .628** .828 Usage Intentions 5.463 (1.247) .950 .969 -.372** .453** .528** .954 A Diagonal cells indicate an A V E (Average Variance Extracted) of the corresponding construct. B Other cells indicate inter-construct correlations. Table 3.5 Descriptive Statistics High Emotion Group Low Emotion Group RA WAD RA E B A RA WAD RA E B A Sub total t-test result RA WAD RA E B A Sub total t-test result Control Mean 4.72 5.54 5.13 t=-3.06; D=.004 5.01 5.44 5.22 t=-1.59; p=.118 4.86 5.49 SD .99 .91 1.03 1.09 .78 .96 1.04 .84 Effort Mean 2.37 2.17 2.27 t=.79; p=.432 2.43 2.29 2.36 t=.60; p=.546 2.40 2.23 SD 1.05 .69 .88 .79 .82 .80 .92 .76 Recomm. Quality Mean 4.42 5.74 5.08 t=-5.36; D=.000 5.12 5.64 5.38 t=-2.54; D=.014 4.77 5.69 SD 1.01 .68 1.08 .79 .63 .76 .97 .65 Usage Intentions Mean 4.86 6.01 5.44 t=-3.47; D=.001 5.40 5.57 5.48 t=-.50; p=.619 5.13 5.79 S D 1.43 .80 1.29 1.26 1.17 1.21 1.36 1.02 66 Table 3.6 ANCOVA Results: Perceived Control Source Sum of Squares df Mean Square F Sig. Corrected Model 12.055(a) 8 1.507 1.598 .136 Intercept 36.664 1 36.664 38.879 .000 Gender .366 1 .366 .388 .535 Previous web experience .477 1 .477 .506 .479 Current web usage frequency .141 1 .141 .149 .700 Tendency to pursue accuracy over effort .284 1 .284 .302 .584 Product Involvement .001 1 .001 .001 .981 Task Emotionality .317 1 •317: .336 .563 RAType 8.929 1 8.929 9.469 .003** Task Emotionality * RA Type .923 1 .923 .978 .325 Error 85.816 91 .943 Total 2781.111 100 Corrected Total 97.871 99 (a) R 2 = .123 (Adjusted R 2 = .046); * indicates p <05; ** indicates p <.01 Table 3.7 ANCOVA Results: Recommendation Quality Source Sum of Squares df Mean Square F Sig. Corrected Model 30.164<a) 8 3.770 5.861 .000 Intercept 45.863 1 45.863 71.286 .000 Gender .272 1 .272 .423 .517 Previous web experience .985 1 .985 1.530 .219 Current web usage frequency .006 1 .006 .009 .926 Tendency to pursue accuracy over effort .413 1 .413 .642 .425 Product Involvement .946 1 .946 1.470 .229 Task Emotionality 3.150 1 3.150, 4.896 .029* RAType 22.740 1 22.740 35.346 .000** Task Emotionality * RA Type 4.294 1 4.294 . 6.675 .011* Error 58.546 91 .643 Total 2824.000 100 Corrected Total 88.710 99 (a) R 2 = .340 (Adjusted R 2 = .282); * indicates p <.05; ** indicates p <.01 67 Table 3.8 ANCOVA Results: Perceived Effort Var i ab les S u m of S q u a r e s df M e a n S q u a r e F S i g . Co r rec ted M o d e l 2 3 2 3 ( a ) 8 .290 .385 .926 Intercept 4 .143 1 4 .143 5 .493 .021 G e n d e r .324 1 .324 .430 .514 P rev ious w e b expe r i ence .158 1 .158 .210 .648 Cur rent w e b u s a g e f requency .042 1 .042 .055 .815 T e n d e n c y to pursue a c c u r a c y ove r effort .166 1 .166 .220 .640 Produc t Involvement .471 1 .471 .625 .431 T a s k Emot ional i ty .083 1 .083 , 1 1 0 ; . 7 4 1 R A T y p e 1.071 1 1.071 1.420 .236 T a s k Emot ional i ty * R A T y p e .057 1 .057 .076 .783 Error 68 .629 91 .754 Total 606 .875 100 Cor rec ted Tota l 70 .953 99 (a) R 2 = .033 (Adjusted R 2 = -.052); * ndicatesp <05; ** indicates p <.01 Table 3.9 ANCOVA Results: Usage Intentions S o u r c e S u m of S q u a r e s df M e a n S q u a r e F S i g . Co r rec ted M o d e l 2 7 . 8 8 9 < a ) 8 3 .486 2 .516 .016 Intercept 64 .974 1 64 .974 4 6 . 8 9 3 .000 G e n d e r .214 1 .214 .154 .695 P rev ious w e b expe r i ence .541 1 .541 .391 .534 Cur ren t w e b u s a g e f requency 4 .734 1 4 .734 3.417 .068 T e n d e n c y to pursue a c c u r a c y over effort 5 .969 1 5 .969 4 .308 .041* Produc t Involvement .222 1 .222 .160 .690 T a s k Emot ional i ty .002 1 .002 . .001 .973 R A T y p e 9 .845 j 1 9 .845 7 .105 .009** T a s k Emot ional i ty * R A T y p e 7.977 1 7.977 5 .757 •018* Error 126 .088 91 1.386 Tota l 3138 .778 100 Cor rec ted Tota l 153 .977 99 (a) R 2 = .181 (Adjusted R 2 = .109); * indicates p <05; ** indicates p <.01 68 Table 3.10 Factor Loadings and Cross Loadings Recommendation Quality (RecQual) Usage Intentions (INT) Perceived Control (CTR) Perceived Effort (EFF) CTR1 0.50 0.37 0.85 -0.30 CTR2 0.62 0.43 0.88 -0.46 CTR3 0.57 0.40 0.93 -0.46 RecQuaM 0.85 0.38 0.61 -0.35 RecQual2 0.89 0.41 0.61 -0.42 RecQual3 0.83 0.40 0.49 -0.47 RecQuaW 0.72 0.53 0.39 -0.26 EFF1 -0.34 -0.27 -0.37 0.81 EFF2 -0.35 -0.32 -0.31 0.80 EFF3 -0.40 -0.35 -0.44 0.80 EFF4 -0.27 -0.22 -0.36 0.68 INT1 0.50 0.96 0.42 -0.37 INT2 0.51 0.97 0.45 -0.38 INT3 0.47 0.93 0.43 -0.39 69 Figure 3.1 Feedback-Control Loop (Morris and Marshall 2004) Figure 3.2 Two Stages of RAs Two Stages of RAs: Users' Tasks in Each Stage: • Construct their preferences • Specify the constructed preferences Preference- Recommendation-Elicitation Stage W Presentation Stage • Review recommended alternatives • If satisfied: make the final choice • Else: return to the previous stage 70 Figure 3.3 RA-WAD Y o u c u r r e n t l y h a v e 100 o u t o f 100 i m p o r t a n c e p o i n t s t o a s s i g n . Please c h o o s e an at tr ibute to indicate its i m p o r t a n c e . A You may skip attributes that you consider unimportant to your purchase. @ Engine breakdown rate View details of engine breakdown rate <@ Brakes breakdown rate View details of brakes breakdown rate © Transmission breakdown rate View details of tranmission breakdown rate <§J Cooling system breakdown rate View details of cooling system breakdown rate © Fuel system breakdown rate View details of fuel system breakdown rate ® Exhaust system breakdown rate View details of exhaust system breakdown rate © Occupant Survival rate View details of occupant survival rate ® Safety features View details of safety features ® Price Range View details of price range Y o u c u r r e n t l y h a v e 100 o u t o f 100 i m p o r t a n c e p o i n t s t o a s s i g n . j P lease ass ign the m a x i m u m of 100. points accord ing to how important t ranmiss ion b r e a k d o w n rate is to y o u . & Type " 0 " , I fyou consider this attribute unimportant. View cars sorted by tranmission breakdown rate View details of tranmission breakdown rate ?V Note that vehicles with low transmission breakdown rates are only a few and cost more. Tranmission breakdown rate ranges from: 'Excellent (5% or lower breakdown rate within the next few years),' 'Good (5%-20% breakdown rate),' 'Average (20%-35% breakdown rate),' 'Mediocre (35%-50% breakdown rate) / 'Poor (50% or higher breakdown rate).' I would consider tranmission breakdown rate when purchasing a used car Decisional Guidance with ( ) out of 100 importance points. [submit] (CanceTj 71 Figure 3.4 RA-EBA Y o u c u r r e n t l y h a v e 6 0 cars in y o u r r e c o m m e n d a t i o n se t . Please choose an attribute to specify the least acceptable level. u m a y s k i p a t t r i b u t e s t h a t y o u c o n s i d e r u n i m p o r t a n t t o y o u r p u r c h a s e . (& E n g i n e b r e a k d o w n r a t e V i e w d e t a i l s o f e n g i n e b r e a k d o w n r a t e © B r a k e s b r e a k d o w n r a t e V i e w d e t a i l s o f b r a k e s b r e a k d o w n r a t e <t) T r a n s m i s s i o n b r e a k d o w n r a t e V i e w d e t a i l s o f t r a n m i s s i o n b r e a k d o w n r a t e <$) C o o l i n g s y s t e m b r e a k d o w n r a t e V i e w d e t a i l s o f c o o l i n g s y s t e m b r e a k d o w n r a t e @ F u e l s y s t e m b r e a k d o w n r a t e V i e w d e t a i l s o f f u e l s y s t e m b r e a k d o w n r a t e © E x h a u s t s y s t e m b r e a k d o w n r a t e V i e w d e t a i l s o f e x h a u s t s y s t e m b r e a k d o w n r a t e © O c c u p a n t s u r v i v a l r a t e V i e w d e t a i l s o f o c c u p a n t s u r v i v a l r a t e © S a f e t y f e a t u r e s V i e w d e t a i l s o f s a f e t y f e a t u r e s © P r i c e R a n g e V i e w d e t a i l s o f p r i c e r a n g e Y o u c u r r e n t l y h a v e 6 0 cars in y o u r r e c o m m e n d a t i o n se t . Please indicate the least acceptable level for tranmission breakdown rate. * V i e w c a r s s o r t e d b y t r a n m i s s i o n b r e a k d o w n r a t e V i e w d e t a i l s o f t r a n m i s s i o n b r e a k d o w n r a t e [ & N o te t h a t v e h i c l e s w i t h l o w t r a n s m i s s i o n b r e a k d o w n r a t e s a r e o n l y a f e w a n d c o s t m o r e . Decisional Guidance • E x c e l l e n t ( 5 ° / o o r l o w e r b r e a k d o w n r a t e w i t h i n t h e n e x t f e w y e a r s ) 2 c a r s f a l l i n t o t h i s l e v e l . • G o o d ( 5 < V b - 2 0 < V o b r e a k d o w n r a t e w i t h i n t h e n e x t f e w y e a r s ) 1 6 c a r s f a l l i n t o t h i s l e v e l . • A v e r a g e f 2 0 < M » - 3 5 < y f a b r e a k d o w n r a t e w i t h i n t h e n e x t f e w y e a r s ) 4 0 c a r s f a l l i n t o t h i s l e v e l . • M e d i o c r e O S ' V b - S O ' W a b r e a k d o w n r a t e w i t h i n t h e n e x t f e w y e a r s ) 5 6 c a r s f a l l i n t o t h i s l e v e l . • P o o r ( S O ' M i o r h i g h e r b r e a k d o w n r a t e w i t h i n t h e n e x t f e w y e a r s ) 6 0 c a r s f a l l i n t o t h i s l e v e l . 72 Figure 3.5 Perceived Control 7 6 5 4 3 2 1 0 5.54, 4.72 5.44 5.01 • • - - RA-WAD - • — R A - E B A High Emotion Task Low Emotion Task Figure 3.6 Recommendation Quality 7 6 5 4 3 2 1 0 5.74 4.42 ^ 5 . 6 4 • 5.12 • • - - RA-WAD - * — R A - E B A High Emotion Task Low Emotion Task Figure 3.7 Perceived Effort 7 6 5 4 3 2 1 0 2.37 2.17 1 - i 2.43 2.29 • • - - RA-WAD - • — R A - E B A High Emotion Task Low Emotion Task Figure 3.8 Usage Intentions 7 6 5 4 3 2 1 -\ 0 6.01 4.86 • • - - RA-WAD HU— RA-EBA High Emotion Task Low Emotion Task Figure 3.9 Structural Equation Model Analyses 75 CHAPTER 4: EFFECTS OF COMPATIBILITY ON CONSUMERS' IN-STORE DECISION MAKING WITH MOBILE RECOMMENDATION AGENTS 4.1 I N T R O D U C T I O N Mobile RAs, the RAs implemented for mobile devices, such as cellular phones and/or Personal Digital Assistants (PDA), have recently attracted the attention of both academics and practitioners (Joshi 2000; Miller et al. 2003; O'Hara and Perry 2001; Van der Heijden 2005; Van der Heijden and Sorensen 2002; Van der Heijden and Sorensen 2005). Mobile RAs are devised to support consumers "en route," including when they are inside a retail store and attempt to make a product choice (Van der Heijden 2005; Van der Heijden and Sorensen 2002; Van der Heijden and Sorensen 2005). Such in-store use of RAs has become possible due to the combination of three trends: (1) the increase in the processing power of mobile devices; (2) the increasingly widespread adoption of mobile devices by consumers; and (3) the increase in the availability of product-descriptions over wireless Internet (Van der Heijden 2005; Van der Heijden and Sorensen 2002; Van der Heijden and Sorensen 2005). Early prototypes of such mobile RAs are the Pocket Bargainfinder and the Shopper's Eye. Mobile RAs at the point of purchase offer consumers significant potential benefits by providing instant access to product information and advice on site, thereby rendering the decision-making process more effective and efficient (Lee and Benbasat 2003; Lee and Benbasat 2004). Consumers lack the resources to perform comparison shopping at the point of purchase (Miller et al. 2003; O'Hara and Perry 2001). Mobile RAs can serve this immediate need, providing web-based reviews and information about products. In addition, mobile RAs reassure the consumer about product quality when s/he wants a second opinion to verify whether or not his/her choice is optimal (Van der Heijden 2005; 76 Van der Heijden and Sorensen 2002). In particular, when the mobile R A is operated by an independent third-party information provider, consumers trust the information offered by the R A more than that offered by salespeople at the store (Komiak, Wang and Benbasat 2004/2005). In Miller et al. (2003)'s survey, 90% of the mobile R A users surveyed said that they found recommendations on a wireless device useful. Nonetheless, very few researchers, except Van der Heijden (2002, 2005a, 2005b), have empirically examined in-store use of mobile RAs from the perspective of consumers' purchase decisions, and considered such issues as the advantages and disadvantages of using mobile RAs for consumers' in-store decision making. Although Van der Heijden was the first to delve into this topic, his studies were limited in the following two respects: First, the RAs he studied were not fully interactive. These RAs lacked an interactive preference-elicitation method and required research assistants to enter consumer preferences for products into the RA ' s algorithms manually. Without an interactive preference-elicitation method, participants cannot modify their preferences once they have entered a retail store. Therefore, the lack of an interactive preference-elicitation method prevents one from investigating consumer' decision making that adapts dynamically to the various factors present within an in-store context. Second, and more importantly, Van der Heijden focused exclusively on the accuracy of decision outcomes as affected by the use of mobile RAs 6 . Other important aspects of mobile R A use, such as its effects on consumers' decision-making efforts and attitudes towards the store, have not yet been fully explicated. Therefore, my first goal in this study was to investigate the effects of mobile R A use, as compared to non-use, on in-store consumers' effort to make decisions and accuracy of decision outcomes. In addition, I contrasted intentions to return to the store 6 Van der Heijden and Sorensen (2005) compared hedonic and utilitarian ratings of mobile RAs as opposed to a plain mobile website. However, the difference was non significant. 77 among these consumers. The focus of this research was an investigation into in-store decision making; thus, it is worthwhile to explore how consumers transfer their attitudes towards the mobile RAs to the store where they use mobile RAs. In addition, I examine the effects of compatibility between the store's product displays and the RA's guidance directions on consumers' in-store decision making. Compatibility occurs when stimuli are paired with a response mode and both stimuli and response mode have a similar format (Kunde 2003). Guidance directions refer to the order in which the RAs guide consumers' decision making. More specifically, it is concerned with the way each R A guides consumers' information acquisition, elicits their preferences for products, and generates recommendations. The issue of compatibility becomes relevant and important to mobile R A use in in-store contexts because R A users conduct two tasks simultaneously: (1) they examine products displayed on the shelves in the store and (2) they utilize the guidance provided by the mobile RAs. Compatibility therefore occurs due to the match between the store's product displays and the RA's guidance directions. Compatibility is an important influence on consumers' decision performance (Chernev 2004; Nowlis and Simonson 1997; Slovic et al. 1990). Compatibility improves decision makers' performance because it exempts them from the additional and error-prone mental operations necessary to align incompatible input and output modes (Slovic et al. 1990). Therefore, if the in-store display and the RA ' s guidance direction are incompatible, a consumer must convert the manner in which information is collected in the store into the way in which the R A guides his/her decision making. Conversely, when there is compatibility, the consumer can utilize the R A instantly without having to reprocess it. Products in a store are usually displayed randomly by alternative, or sorted only by a few attributes, such as price or brand, but not by every attribute simultaneously. There 78 are clear limitations to organizing complex, durable products by every attribute in a store, given that such products have multiple heterogeneous attributes, with various conflicting levels. For instance, most computers have more than 40 attributes (according to the product descriptions found at www.dell.com), some of which are conflicting. More specifically, consider two laptops, A and B: Laptop A has a larger L C D screen size and a smaller storage capacity than laptop B. In this case, it becomes impossible for a store manager to sort the two laptops in an ascending order by both attributes. However, many stationary RAs employ an attribute-driven approach, which presents information about attributes and elicits consumers' preferences for those attributes, rather than an alternative-driven approach, which presents information about alternatives and elicits consumers' preferences for alternatives. This is because consumers find an attribute-driven approach easier to use, less effortful, and more intuitive than an alternative-driven RA, especially for complex products with a number of attributes and alternatives (Payne et al. 1993). However, given that products in a store are displayed by alternative (as discussed above), an attribute-driven R A may not be compatible with in-store product displays. Because of this incompatibility, consumers must expend additional mental effort to align the information gathered from the store and the RA, a process which hampers their decision making. Conversely, an alternative-driven R A is more compatible with in-store product displays, and therefore more likely to enhance consumers' decision making. The second goal of this study is to investigate whether or not compatibility between RA' s guidance and the store's product displays affects decision accuracy, decision-making effort, perceived control and finally intentions to use the RAs. To accomplish the second goal, I designed and compared two mobile RAs, an ALternative-driven R A (RA-AL) and an ATtribute-driven R A (RA-AT). R A - A L and RA-AT are based on two opposite guidance directions: i.e., attribute- and alternative-driven approach. 79 The remainder of this chapter is organized as follows. In Section 4.2, I develop hypotheses about how consumers who use RAs for in-store decision making differ from those who do not in terms of decision-making processes and outcomes. Then, the distinctions between RA-AT and R A - A L are illustrated, particularly with regards to their compatibility with in-store product displays, and hypotheses about the effects of compatibility on consumers' decision-making are developed. Section 4.3 presents the research method utilized to empirically compare first the control group and both R A groups, and second, the two R A groups, i.e., R A - A L and RA-AT. Section 4.4 presents data analyses and their results. Chi-square results to compare the accuracy of decision outcomes between the different groups are presented, followed by A N C O V A results that contrast the two types of RAs. Finally, Section 4.5 discusses the results of the experiment and their implications for theory and practice. 4.2 T H E O R Y D E V E L O P M E N T 4.2.1 Use of Mobile RAs for In-Store Decision Making In this section, I investigate whether or not mobile R A use increases accuracy of decision outcomes, decreases consumers' perceived effort to make decisions, and increases their intentions to return to the store. The other two dependent variables included in the comparisons of R A - A L and RA-AT in the subsequent section - perceived control and usage intentions - are not included, because they are relevant only to R A users but not to non-RA users. R A use increases the accuracy of consumers' purchase decisions by providing enhanced search and comparison features (Xiao and Benbasat 2007). Specifically, certain types of RAs help consumers to compare products in-depth before making a final decision (Haubl and Trifts 2000). Using comparison features, consumers can identify suboptimal alternatives and choose the best alternative. 80 I argue that search and comparison features also increase the quality of in-store decision making based on Russo's notion that these features transform "available" information into "processable" information (1977). In a retail store without RAs, product information is provided in a much less systematic way, usually via product price tags (i.e., the tags that show product price and brief information, which are attached to the display shelves). Information about certain attributes, such as a computer graphics card, is frequently shown on one product's tag while missing on another. This makes it extremely difficult for consumers to collect, integrate, and comprehend information about different products (Russo, Staelin, Nolan, Russell and Metcalf 1986). Moreover, even when information about all product attributes is consistently available, consumers often fail to utilize that information unless it becomes "processable" (Russo et al. 1986). In his classic study, Russo (1986) examined the use of unit price information by supermarket shoppers, and found that usage increased when the information was brought together in the form of organized lists ranking available brands by increasing unit price. Russo argued that standard presentations of unit price information, (for example, posted on the supermarket shelf under each item), made prices difficult to compare. Russo's ranked list made the available unit price information much more accessible, and was therefore more likely to be used by consumers. Simply making information available is not sufficient; information must also be easily "processable," and presented in an accessible form(s). R A search and comparison features make information processable rather than simply available, so that consumers can utilize the information more readily and consequently make more accurate decisions. Therefore, one can predict that mobile R A users will make more accurate decisions than unaided consumers. Prior to further developing my hypotheses on decision accuracy, it is important to clarify that, whereas in the previous chapter the focus was on the subjective quality of consumer decisions, in this experiment I focus on the objective accuracy of decisions, 81 because the latter better supports whether or not R A use will help consumers reach accurate decisions in a retail store. "Decision accuracy" is measured by examining whether a product chosen by a consumer is a non-dominated (an optimal decision) or dominated (a suboptimal decision) alternative (Diehl 2003; Haubl and Trifts 2000; Swaminathan 2002), while "subjective quality of a consumer's purchase decision" is measured by investigating the level of a consumer's confidence in the RA ' s recommendations (Haubl and Trifts 2000; Swaminathan 2002). Subjective quality, however, does not always reflect the objective accuracy of a decision (Kotteman, Davis and Remus 1994; Olson and Widing 2002). According to Kotteman et al. (1994), DSS users often misjudge their performance. DSS users systematically overestimate how much DSS helps their decision making. However, this subjective belief in their decision performance does not parallel the objective accuracy of the decisions made (Kotteman, Davis and Remus 1994). In Kotteman et al.'s experiment (1994), only half of DSS users made more accurate decisions, while the other half made decisions that were inferior to those made by unaided users. For this reason, I chose objective decision accuracy over subjective quality of decisions for the comparison between mobile R A users and non-users, and developed my hypothesis on the difference in objective accuracy of decisions. HI. In a retail store, mobile RA users will make more accurate decisions than non-users. In purely computer-based environments such as online shopping situations, the application of RAs is known to reduce the amount of effort exerted during decision-making processes (Xiao and Benbasat 2007). A wealth of empirical data has been collected showing that R A users spend significantly less time searching for information and completing shopping tasks, compared to non-users (Hostler, Yoon and Guimaraes 2005; Vijayasarathy and Jones 2001). Based on this empirical evidence, one can 82 confidently conclude that R A use reduces the effort required by consumers making decisions online or in web-based storefronts. However, to date there has been no research into whether or not R A use reduces the effort required to reach a product choice in a retail store. On the one hand, it could be argued that using an R A constitutes an additional task for consumers: R A users in a retail store are not only evaluating the store itself and its products, but also looking up the product information provided by the R A on the mobile website(s). In other words, R A users need to multi-task (i.e., simultaneously perform an in-store search and an online search with the RA), whereas R A non-users conduct only one task (i.e., an in-store search). Thus, the additional search capabilities and extra information provided by the RAs may add to the cognitive load of consumers, thereby increasing their decision-making effort. In a similar vein, Benbasat and Todd (1996) contended that use of RAs is not cost-free: utilization of the RA' s features requires additional mental effort. On the other hand, mobile RAs simplify a tedious process, screening and sorting products based on consumers' expressed preferences, thereby reducing consumers' information search and enabling them to focus on the alternatives that best match their specific preferences (Xiao and Benbasat 2007). In addition, whereas R A non-users make decisions entirely on their own, R A users can delegate their decisions to the RA, thereby sharing the stress involved in making a choice (Komiak and Benbasat 2006). In particular, when the mobile RAs are operated by a third-party rather than the retail store, consumers feel more willing to delegate their decision(s) to the agent, safe in the knowledge that the agent's recommendations are free of the inevitable bias of the salesperson and the store (Campbell and Kirmani 2000). This voluntary delegation reduces the stress and anxiety involved in making choices alone, and hence further reduces consumers' perceived effort level. Consequently, I posit that the effort saved by using mobile RAs will exceed the effort required to use the RA. Therefore, 83 H2. In a retail store, mobile R A users will report less perceived effort than non-users. Because the use of mobile RAs increases the accuracy of decisions and reduces the effort required to make those decisions, mobile R A users will form positive attitudes towards the mobile RAs. Consumers will transfer their positive attitudes towards an additional event when they perceive associations between the original event and the extension (Dickinson and Barker 2007; Murdock 1985; Shimp, Stuart and Engle 1991). In this study, the mobile RAs and retail stores complement each other. In particular, consumers review product information and get recommendations from the mobile RAs while actually examining products in the store. In other words, the physical store complements the mobile RAs in the sense that the store enables users to visually examine products. Hence, consumers will develop natural associations between the R A and the store, and then transfer their positive attitudes toward mobile RAs to the retail store. Consumers' attitudes include co-native components (actions/behaviours, such as intentions to return to the store) in addition to a cognitive component (beliefs) and an affective component (feelings) (Dickinson and Barker 2007). As a result of transferring their positive attitudes to the retail store, mobile R A users are more likely to return to the store than R A non-users who have not built up the same degree of positive attitudes towards the store. Therefore, H3. Mobile RA users will show higher intentions to return to a retail store than R A non-users. 4.2.2 Alternative-Driven RA vs. Attribute-Driven RA In this section, I compare two RAs - attribute-based R A (RA-AT) and alternative-based R A (RA-AL) - in terms of decision accuracy, perceived effort, perceived control, and usage intentions. "Attribute-driven (AT)" and "alternative-driven (AL)" indicate the order in which the RAs guide consumers' decision making. More specifically, it is 84 concerned with the way each R A guides consumers' information acquisition, elicits their preferences for products, and generates recommendations. R A - A L prompts consumers to acquire information about alternatives, and elicits their preferences for alternatives, then generates recommendations based on those preferences. Figure 4.1 illustrates how R A - A L works in a diagram and Figure 4.2 shows R A - A L screenshots. R A - A L first presents a list of alternatives7 (see the "product-list page" in Figures 4.1 and 4.2), from which users can view the details of alternatives ("detailed-information-about-product page"). While doing this, users can also review attribute details, such as attribute definitions and range of attribute levels ("information-about-attribute page"). R A - A L then allows consumers to compare alternatives pair-wise ("comparison page"), narrowing their search to an ever-decreasing number of products. Next, R A - A L recommends products similar to the alternatives favoured and therefore saved for further consideration from the pair-wise comparisons ("recommendation page"). Finally, R A - A L enables users to compare their favoured alternatives selected from the pair-wise comparisons and the R A - A L ' s recommendations until they reach their final choice. Examples of R A - A L can be found at www.bestbuy.com. — INSERT [Figure 4.1 and 4.2] H E R E — Conversely, RA-AT prompts consumers to acquire information about attributes and elicits their preferences for attributes, then provides recommendations based on those preferences (see Figures 4.3 and 4.4). Specifically, RA-AT provides a list of product attributes (see the "attribute-list page" in Figures 4.3 and 4.4), from which they can review the details of the attributes ("information-about-attribute page"). Then, RA-AT asks consumers to specify their preferences for the relevant attributes and to indicate the 7 Although R A - A L listed products alphabetically, products in the store were displayed randomly. A perfect match in both orders was avoided as this could have increased the artificiality of the experiment. Given current technologies, RAs cannot adjust the order of their product lists according to the store's product displays, especially when the R A is a third-party. . 8 5 importance of the chosen preferences ("preference-specification page"). Finally, RA-AT provides recommendations that satisfy the consumers' expressed attribute preferences ("recommendation page"), and the consumers can then view the details of recommended products ("detailed-information-about-product page"). Examples of this type of R A can be found at www.ActiveBuyers.com and www.mvProductAdvisor.com. — INSERT [Figure 4.3 and 4.4] H E R E ---Unlike RA-WAD and R A - E B A described in Chapter 3, R A - A L and RA-AT are loosely coupled with decision strategies. In other words, RA-WAD and R A - E B A both compel consumers to use a particular dedicated strategy, i.e., WAD and E B A , respectively. Conversely, R A - A L and RA-AT offer consumers interface features with which they can employ various strategies at their own discretion. While using the pair-wise comparison feature of R A - A L , a consumer can apply either ADDIF and/or M C D strategies, the two compensatory strategies utilized for pair-wise comparisons, (See section 2.1, Decision Strategies, for detailed discussions; Bettman et al. 1993). However, s/he could use a less accurate and non-compensatory strategy, such as SAT (see section 2.1 for a detailed description of this strategy). This is because using the pair-wise comparison feature does not guarantee that the consumer actually compares the two products added to the comparison-page. S/he might use SAT by simply evaluating one product at a time: S/he first evaluates the product on the left, then the product on the right, and repeats this process until s/he reaches the final choice. Finally, i f the user chooses to view the recommendations made by R A - A L , R A - A L applies EQW (Payne et al. 1993). Specifically, when R A - A L finds the alternatives with values similar to those of the user-chosen product in terms of every attribute, R A - A L ignores the relative importance of attributes and assumes equal weight for each attribute (Bettman et al. 1993). RA-AT users can employ both E B A and simplified WAD strategies. E B A and 86 WAD are integrated in RA-AT because they are widely accepted and employed by many attribute-driven RAs (Haubl and Trifts 2000). The two strategies are integrated as follows: RA-AT requires users to indicate first their preferred attribute levels and then the importance of these attributes in their purchases. There are three importance levels: (1) extremely important, (2) important, and (3) somewhat important. If consumers choose the level "extremely important8," then any alternatives that do not meet this level are eliminated, which means that the E B A strategy is applied. When consumers choose lower importance levels, such as "important" and "somewhat important," RA-AT applies a simplified WAD to calculate an overall score for each alternative9. In sum, R A - A L and RA-AT provide users with interface features which allow them to select and utilize decision strategies at their own discretion. This is a significant difference compared to RA-WAD and R A - E B A , which require consumers to use particular strategies. 4.2.3 Effects of Guidance Directions on Decision Processes Prior studies have shown that information presentation formats, both attribute- and alternative-driven, influence the way in which consumers acquire product information (Jarvenpaa 1989). Consumers tend to reduce their effort to process information when they are exposed to new information, acquiring but not adapting the new information (Payne 8 During the experiment, participants were made fully aware that the choice of "extremely important" would result in the elimination of alternatives that did not meet the chosen level, whereas the choice of "important/somewhat important" would not have this effect. 9 RA-AT first attains the gap score between the user-preferred attribute level and the attribute level of each alternative. For instance, if a user chooses level 3 ($101-$200) out of 4 price levels, and product A has a level 4 price, then the gap score is 1 (4-3). Then, RA-AT multiplies the gap score by the standard score of the user-chosen importance level (e.g., 5 = somewhat important, 10 = important). Therefore, if the user chooses "important," then the multiplied score would be 10 (1*10). RA-AT adds up the gap score for each attribute to attain the overall gap score for each alternative. RA-AT then recommends the alternatives with the lowest gap scores (i.e., the alternatives that best match the user-specified levels) to the consumer. I chose to simplify WAD in this particular way because of the limited input-output devices of mobile computers. Many consumers find it cumbersome to enter numbers using the small keyboard or keypad on mobile devices (Pascoe, Ryan and Morse 1998; Pascoe, Ryan and Morse 2000). Also, the provision of more importance options, such as a 7 point scale, was not feasible, as the small screen of a mobile device did not offer enough space. 87 et al. 1993). This tendency applies equally when an information presentation format is incongruent with task requirements. Jarvenpaa (1989) found that participants used alternative-based processing when provided with a graph depicted by alternatives, even i f the task instructions specifically elicited attribute-based processing. Conversely, participants used attribute-based processing when provided with an attribute graphic format, even though the task required alternative-based processing. This suggests that acquisition direction is a function of information presentation formats rather than specific task structures. RA-AT focuses on attribute-driven processing and elicits consumers' preferences for attributes, prompting consumers to deliberate on individual attributes rather than individual products (Olson and Widing 2002). Conversely, R A - A L explicitly requires consumers to review the details of various products and compare pairs of products until they narrow their search to a few options. As Jarvenpaa (1989) has noted, RA-AT and R A - A L s different guidance directions influence the ways in which consumers acquire information. That is, R A - A L users tend to acquire information about alternatives, such as individual product features, when utilizing the product list with which R A - A L prompts them to begin their search. Conversely, RA-AT users tend to acquire information about attributes, such as attribute definitions, when utilizing the list of attributes presented by RA-AT. As a result, R A - A L users will examine alternatives more often than RA-AT users. H 4 . M o b i l e R A - A L u s e r s w i l l v i e w a l t e r n a t i v e s m o r e o f t e n t h a n m o b i l e R A -A T u s e r s . 4 . 2 . 4 C o m p a t i b i l i t y b e t w e e n t h e R A ' s G u i d a n c e D i r e c t i o n s a n d t h e S t o r e ' s P r o d u c t D i s p l a y s The alternative- and attribute-driven directions of R A - A L and RA-AT differ in terms of compatibility with the way in which products are displayed in a retail store. I argue that R A - A L is more compatible with the store's product displays than RA-AT for 88 the following reasons. In a retail store, there are physical limits to sorting and displaying durable and complex goods on shelves by attribute: First, complex durable goods, such as computers, have a number of common attributes (e.g., graphics/sound cards, CPU, R A M , R O M , HDD, etc.), but there are often conflicts among attributes. For instance, many stores offer computers with more than 40 attributes (e.g., BestBuy), and may sort these computers first by brand, then by price within brand, then by CPU type within price, and so on. However, attribute levels are not always correlated; for example, some computers may have a lower price and a faster CPU speed. Such conflicts make it impossible to sort all the computers systematically by one attribute under another. Furthermore, some attributes of complex products are distinct and hence not common to all products. For instance, some computer models are specially designed for playing media files with enhanced video and audio functions, while other models are equipped with special features for business purposes, such as fingerprint recognition for enhanced security. Such distinct media and security features are not common to all computers, making it impossible to sort all the computers by any one of these attributes. In sum, because complex products have conflicting and non-common attributes, retail stores cannot display or organize such products systematically by attribute, and instead tend to display them randomly product by product. R A - A L is more compatible with alternative-based displays in retail stores, for the following reasons. One can reasonably assume that most (if not all) consumers who visit a retail store want to look around to see what products are available on the shelves. In other words, even though the depth of examination may differ according to the individual consumer's shopping traits and the number of products available in the store, most consumers will naturally examine all or at least some of the products inside the store they have chosen to visit. 89 R A - A L allows users to select alternatives first and then compare the selected alternatives by attributes, while RA-AT requires users first to specify their attribute preferences before presenting alternatives. Thus, whereas R A - A L users can find product information for a product they have seen on display in the store promptly by choosing products from the list organized by alternative, RA-AT users must specify their attribute preferences before they can begin gathering product information about the same product. Therefore, the decision guidance provided by R A - A L most closely resembles the way in which products are displayed in the retail store. In contrast, RA-AT decision guidance requires that consumers transform the product information into attribute information. Consequently, RA-AT is less compatible with the way in which products are displayed in a retail store. 4.2.5 Effects of Compatibility on Decision Accuracy, Perceived Effort, Perceived Control, and Usage Intentions The significance of the compatibility between input and output has long been recognized by researchers of human performance (Chernev 2004; Slovic et al. 1990). Compatibility not only saves users from the additional mental processes required to convert incompatible input (or output) modes into compatible output (or input) modes, but also protects them against any potential errors that might arise because of the incompatibility. Slovic et al. (1990) illustrated the positive effects of compatibility as follows: Engineering psychologists have discovered that responses to visual displays of information, such as an instrument panel, will be faster and more accurate if the response structure is compatible with the arrangement of the stimuli (Fitts and Seeger 1953; Wickens 1984). For example, the response to a pair of lights will be faster and more accurate if the left light is assigned to the left key and the right light to the right key. Similarly, a square array of four burners on a stove is easier to control with a matching square array of knobs than with a linear array. Also, the concept of compatibility has been extended beyond spatial organization. The 90 reaction time to a stimulus light is shorter for a pointing response than for a vocal response, but the vocal response is faster than the pointing response if the stimulus is presented in an auditory mode. It is evident that some input-output configurations are much more compatible than others and therefore yield better performance [Emphasis added]. As argued above, RA-AT is less compatible with a store's product displays than R A - A L . Hence, RA-AT users must exert additional mental effort to resolve the incompatibility between the store's product display and the RA' s guidance direction. In particular, a RA-AT user must specify his/her preferences for attributes in order to check the details of certain products s/he has found interesting in the store. To make matters worse, i f RA-AT does not recommend the product the user wanted to review, the user must revise his/her preferences for attributes until RA-AT recommends that particular product. The ensuing repeated specification of preferences increases the possibility of making errors. In addition, as the complexity of decision tasks increases, consumers tend to use an elimination heuristics rather than a compensatory strategy in an attempt to reduce their effort (Payne et al. 1993). RA-AT users, irritated by the need to revise their preferences repeatedly on several attributes, may grow more intent on focusing on fewer attributes by using the "extremely important" level in the hope of getting to the alternative they would like to observe more quickly; this may lead to a less accurate choice of products (Payne et al. 1993; Todd and Benbasat 1991). Consequently, I posit: H5. Mobile R A - A L users will make more accurate decisions than mobile RA-A T users. In accordance with previous research on the positive effects of compatibility, I argue that consumers' decision making will become much more efficient when using a R A that is compatible with the way consumers shop and gather information inside the store. In contrast, incompatibility between guidance directions and in-store display(s) is 91 likely to slow down consumers' decision-making processes. Therefore, H6. M o b i l e R A - A T u s e r s w i l l p e r c e i v e h i g h e r l e v e l o f e f f o r t t h a n m o b i l e R A -A L u s e r s . R A - A L users can select the product in which they are particularly interested and immediately seek further information by clicking on the name of that product in the list. In contrast, RA-AT users have to go through the preference-elicitation stage during which they specify their preferences for attributes. Furthermore, RA-AT may not recommend the particular product about which the consumer wanted further information about, as noted earlier. Clearly, consumers have more limited control over how they navigate with RA-AT compared to R A - A L . Therefore, H7. M o b i l e R A - A T u s e r s w i l l p e r c e i v e l e s s c o n t r o l t h a n m o b i l e R A - A L u s e r s . RA-AT appears less effective than R A - A L in terms of decision accuracy (H5), perceived effort (H6), and perceived control (H7), the three factors that influence consumers' intentions to use RAs, as argued in Chapter 3. Therefore, one can conclude that consumers will show stronger intentions to use R A - A L than RA-AT. H8. M o b i l e R A - A T u s e r s w i l l s h o w l o w e r u s a g e i n t e n t i o n s t h a n m o b i l e R A -A L u s e r s . 4.3 R E S E A R C H M E T H O D I conducted a laboratory experiment in order to test the stated hypotheses empirically. Like the experiment described in the previous chapter, this experiment allowed close control over independent, moderating, dependent, and possibly confounding variables to achieve a high degree of internal validity (Singleton and Straits 1999). A 3 x 1 factorial design with one between-subject factor was utilized. The between-subject factor had three levels: RA-AT, R A - A L , and the control group. 92 4.3.1 Alternatives and Attributes I chose printers as the focus product for this experiment because they are durable and complex goods with a large number of alternatives and attributes, and participants would therefore have a strong rationale for seeking help from the RAs during their purchase. According to the review of previous research on RAs by Xiao and Benbasat (2007), consumers do not benefit from using RAs when purchasing simple products, as it is easier to process information about simple products without R A assistance. In addition, given that the recruited participants were students, who would presumably be interested in and want to purchase printers for their school work, printers seemed to be a particularly suitable and relevant product category for the experiment. The 22 different printer models chosen for the experiment covered the range of printers actually sold in Canada at the time the experiment was conducted. I investigated all printers sold both on- and offline to make sure all possible printer models were included. By including all possible models sold at that time, I ensured that participants looking for a particular model would not be disappointed by its absence in the store, as this would affect their perceptions of the RAs. Table 4.1 shows 11 printer attributes, chosen from the lists of common attributes mentioned in technology review Websites, such as www, consumerreports .org, www.znet.com, www.cnet.com, and online storefronts such as www.staples.ca and www.bestbuy.ca. The 11 printer attributes are brand, black and white print speed, colour print speed, colour resolution, print quality, total cartridge cost, maximum printable paper width, compatible operating systems, media card support, portability, and price. In the experiment described in Chapter 3, attribute levels were manipulated in order to create attribute conflicts; in this second experiment, however, the ranges and levels of each attribute were drawn from real products. - INSERT [Table 4.1] H E R E -93 As was in Study 1, the RAs provided a link that opened a pop-up window containing detailed description about each attribute so that users who were not familiar with these attributes could learn about each attribute (See the pop-up window in Figures 4.2 and 4.4). 4.3.2 R A Design How R A - A L and RA-AT work conceptually was explained above in Section 4.2.2. This section presents the features of the two RA's interfaces in more detail (see Figures 4.1, 4.2, 4.3, and 4.4). The two RAs employed identical visuals in terms of colours and designs, but guided consumers' decision making in two different ways, described below. Both R A - A L and RA-AT introduction pages contained a welcoming message and a brief introduction to the R A in question. R A - A L presented participants with a list of alternatives. As one page allows only enough space to present 11 alternatives, the range of 22 printers was presented across two pages. On this product-list page, participants proceeded to the page offering detailed information about each printer ("detailed-information-about-product page[s]") by selecting the product name. The detailed-information-about-product pages presented the 11 common printer attributes, as well as additional details. These additional details consisted of other technological specifications, including expert reviews (such as pros, cons, and summaries), dimensions (Width x Height x Depth), cartridge types, network features, and interfaces (e.g., USB or parallel). The expert reviews and detailed specifications were similar to those found on other review websites. Because most of the well-known technology-review websites provide information about these additional features, and the lack of such information would hamper the realism of both the RAs and the experiment, I decided to provide it to the participants. On an information-about-attribute page, participants could click on a "?" button placed next to each attribute in order to review that particular attribute, i f they were not familiar with it. 94 Participants were allowed to decide for themselves whether or not to review the detailed-information-about-product pages, but they had to return to the product-list page in order to add printers to the comparison page. Participants added printers they were particularly interested in to the comparison page by checking a "Compare" box next to each product. After adding all the selected printers, participants clicked on another "Compare" button in the top right corner of the page. This opened a comparison page on which participants were able to contrast selected printers, by attributes, side by side. Because one page contains two products, participants were allowed to move across pages to see all the printers. When comparing two products side by side, participants clicked a "Remove" button to eliminate any products they found inferior following the comparison. Participants were told to continue the pair-wise comparisons until they reached one or two preferred products. Once they had narrowed their choices to one or two final products, they were given the option of clicking a "More Like This" button, which brought up six recommended products similar to the remaining one in terms of all attributes. Participants then reviewed the similar products to verify whether their original remaining choice was indeed the best, or whether there were better printers still available. If they found that one (or more) of the newly recommended printers was better than the original remaining selection, they could click an "Add to Comparison" button to add this new printers to the comparison page. They could then compare the new addition to the original remaining selection. Participants continued to make such comparisons until they reached one final choice. In contrast to R A - A L , following the introduction page, RA-AT provided participants with the list of 11 printer attributes. On this list page, participants clicked a "?" button beside each attribute to review that particular attribute. They then selected attributes they thought important and prompted to specify their preferred level for each attribute and to indicate how important that attribute was on a 3-point-scale ("extremely 95 important," "very important," and "somewhat important"). Next, participants clicked the "View Results" button, and this brought up a list of up to six recommendations, leading to separate recommendation pages. Each recommendation page provided only enough space to present two products; thus six printers were presented across three pages. At this stage, participants could click on the name(s) of any printer(s) they were particularly interested in, and thus bring up the detailed-information-about-product pages presenting the 11 common attributes as well as other additional attributes (these pages functioned in the same way as the detailed-information-about-product pages in R A - A L ) . If they were not satisfied with the recommendations provided, participants could click the "Close" button to return to the attribute-list pages and revise their preferences, and this in turn would bring up new recommendations. 4.3.3 Computing Devices Each participant was provided with Personal Digital Assistant (PDA), specifically the HP iPAQ Pocket PC hi940, which runs on the Pocket PC 2003 operating system and has a 240 x 320 pixel resolution color touch screen. Participants used a stylus to enter input on the PDAs. The PDAs were connected through a built-in Wi-Fi card to a campus-wide wireless Internet connection, which is built using the wireless Ethernet standard Wi-Fi IEEE 802.11b. Because the experiments were conducted during the university's summer break, the usual problems associated with a busy wireless network, i f any, such as interference and instability, were negligible. 4.3.4 Operationalization of the Dependent Variables Six dependent variables were captured: (1) perceived effort, (2) perceived control, and (3) intentions to use the RAs for future shopping, (4) intentions to return to the store, (5) the number of products viewed by participants, and (6) the accuracy of decision outcomes. Table 4.2 lists the dependent variables. -- INSERT [Table 4.2] H E R E -96 Among these, perceived effort, perceived control, and intentions to use the RAs were calculated using the same measures as the experiment described in Chapter 3 (please refer to section 3.3.4 for further details). Intentions to return to the store was measured on a 7-point scale (Table 4.2). Coyle and Thomson's (2001) three items were applied: (1) Assuming that I have access to this store, I would like to visit the store next time I need to shop for a printer; (2) / would like to visit a store similar to this store next time I need to shop for a printer, (3) / would shop at a store that is similar to this store in the future (Coyle and Thomson 2001). The number of times participants viewed alternatives was measured by counting the number of times each participant clicked on the detailed-information-about-the-product pages. I deliberately considered the number of times participants viewed the detailed-information-about-the-product pages instead of the number of products participants included in the comparison page (RA-AL) or the recommendation page (RA-AT), because I judged that participants might not have actually reviewed all the products included in those pages. Indeed, it is likely that participants who did not like the alternatives recommended by RA-AT only briefly skimmed the recommended products, and then returned to the preference-elicitation page to revise their preferences. Similarly, R A - A L users might have included some products in the comparison page accidentally, and therefore quickly dropped those products without carefully reviewing them. Since the number of times alternatives were viewed measure counts the times products were reviewed seriously, and this number indicates the degree to which participants gathered product information carefully, the frequency with which participants actually clicked on the products and reviewed product details provides a more useful figure than the number of printers included in the comparison or recommendation pages. I also measured the "objective" accuracy of decision outcomes. For this purpose, many previous studies have employed the concept of the choice of "non-dominated 97 products" (i.e., products not objectively inferior to any alternative) and measured the share of chosen products that are "non-dominated." In order to measure the share of such non-dominated products, previous research manipulated the attribute levels of all alternatives so that some products were better than others in every attribute. However, because I employed real products, and because attribute conflicts are inherent in any efficient market, the use of non-dominated products was not feasible for this study. For instance, those printers with the highest levels in every attribute are often also the most expensive ones. Given attribute conflicts, it was infeasible to find a product that satisfies every consumer's needs. Some consumers, including light, beginner printer-users, are likely to prefer a base model which is inexpensive and has basic levels of all attributes, while others, including expert printer-users, have well-established preferences for attributes and prefer models with decent performance (Swaminathan 2002). I therefore included in the set of printers two products that best satisfied the particular needs of two segments of consumers: (1) HP PhotoSmart 8050 for light, beginner users who print only occasionally for home use; and (2) Canon Pixma iP5200 for heavy, experienced expert users (See table 4.3). In the experiment, I provided participants, as a part of their decision task, with descriptions of the two types of consumers for whom they were asked to choose printers and asked them to choose a printer for each type of consumers. Then, I measured the frequency with which these two printers were selected as an indicator of decision accuracy. A detailed explanation of why the two printers were selected for the two types of consumers is provided in the section 4.3.6, "Desirable vs. Less Desirable Printers." -- INSERT [Table 4.3] H E R E -4.3.5 Experimental Tasks Participants were given two tasks: (1) choosing a printer as a birthday present for a . 98 family member and (2) selecting a printer as a wedding gift for their best friend. The order of the two tasks was counterbalanced; details of both tasks are shown in Appendices A.4 (decision tasks for the R A groups) and 4.5 (decision tasks for the control group). The family member in the experimental task was described as a novice user who had little experience or knowledge of printers and who would have a printer for occasional home use; the friend was characterized as an experienced user who had substantial experience and knowledge of printers and who would use the printer for his/her work. No definite price limits were set for either of the tasks, because participants given price limits would be encouraged to use the E B A strategy, and thereby be unintentionally driven to favour RA-AT. However, a complete lack of price constraints would remove any incentive for participants to strive to save money on their purchases, and lead them to choose only expensive models with desirable attribute levels. In order to avoid this problem, I required participants to begin by considering what price range (not limit) would be appropriate for each task, and then to examine whether or not their chosen printers had stayed within this budget range upon completion of both tasks. 4.3.6 Desirable vs. Less Desirable Printers This section explains why the HP PhotoSmart 8050 and the Canon Pixma iP5200 were deemed the best printers for the two types of printer users (a light beginner user and a heavy expert user), based upon (1) the attribute levels and (2) reviews by consumer magazines. Table 4.4 shows all the attribute levels, except operating systems, for each printer. Given that all 22 printers support the two major operating systems (Windows and Mac OS), operating systems are not shown in Table 4.4. - INSERT [Table 4.4] H E R E -99 Two products, the Canon Pixma IP5200 and the HP Photosmart 8050, showed better values in many attributes than other products in the same price range ($101 -$200); indeed, they even surpassed more expensive products. The Canon Pixma iP5200 has the highest total scores (20 points) that add up all attribute levels; the HP Photosmart 8050 has the second highest (18 points). The highest total scores indicate that these two products are better than other printers in many aspects. More specifically, the Canon Pixma iP5200 was better than all the other printers (including more expensive printers) in terms of maximum colour resolution, print quality, and maximum black and white print speed. Also, it had the second highest levels (3 out of 4) for cartridge cost. Thus, the Canon excelled in such attributes as quality, speed and cartridge cost, considered the most important by heavy, expert users (Consumer Reports, July 2007). The attributes for which the Pixma had the lowest levels (i.e., 1 out of 2 for portability and availability of media card slots) were not critical to the expert users described in the experimental task given to participants. The July 2007 issue of Consumer Reports selected the Canon Pixma iP5200 as "the best all-around performer," noting that, "The Canon (Pixma iP5200) stands out for quality, low print costs, and speed. Photos and text were excellent." The HP Photosmart 8050 was superior to the other printers in terms of maximum colour print speed. In addition, it had the second highest levels (3 out of 4) in terms of cartridge cost and print quality. It had medium but not low levels for most other attributes, including maximum colour resolution. It scored on the lowest level for portability, which, does not matter to the novice user described in the task. Because of its excellent speed, above-average quality, and all-average attributes at a relatively low price ($125.46), PC Magazine (March 2006) placed the HP Photosmart 8050 among the top five photo printers of the year, along with the Canon Pixma iP5200. In addition, MacWorld (November 2005) reviewed the Photosmart as the best "base model," stating that "The differences between a top-of-the-line photo printer and a base model - aside 100 from the purchase price - are often just special features that you may or may not want to use. For example, the inexpensive HP Photosmart 8050 may not feature some of the niceties that more expensive models include-•• but it does do a really good job at printing photos - a lot of the time, that's all you need" (November 2005). In sum, the Canon Pixma iP5200 was a desirable choice for the task involving the friend who was an expert printer user, while the HP Photosmart 8050 was a good choice for the family member who was a novice user. 4.3.7 Store Design Figure 4.5 shows the printer retail store that was created inside the laboratory. The store was approximately 3.5 square meters in size. The product shelves were placed against each of the three walls (there were windows in one of the walls where shelves could not be installed). There was one door through which participants entered and exited the store. Instead of bringing the selected printer to the cash register as in a real retail store, participants indicated their final purchase choices by writing down the name of the printer they had selected on a notepad on the desk before exiting the store. - INSERT [Figure 4.5] H E R E -The store was made to appear as realistic as possible for the participants. The pictures of 22 printers were displayed on the shelves within the store. Because of a number of issues related to budget, logistics, and return of the products used for the experiment, I chose to display life-size pictures of printers, rather than real printers (see Figure 4.6). The use of pictures instead of real products is widely accepted and implemented in other studies investigating in-store purchase choices (for example, see Van Der Heijden 2005). According to the classification of product attributes in Suh and Lee (2005), the salient attributes of printers are search attributes (such as printing speed, resolution, quality, and toner price) rather than experiential attributes (such as size, look, 101 and weight). Search attributes can be appropriately described without direct sampling of the product, and are sometimes better illustrated through the use of indirect media such as brochures or consumer reviews (Suh and Lee 2005). Given the salient features of printers are rather search attributes, the use of pictures instead of real printers would not hamper creating shopping environment in a realistic manner especially in a situation where a provision of real printers is infeasible. - INSERT [Figure 4.6] H E R E -The printer pictures were randomly arranged to mimic the unsystematic style of in-store product displays. In order to achieve as realistic a product display as possible, my arrangement closely emulated the displays at Staples. Not only is Staples one of the largest electronics retail chains in North America, but it also has a branch on the University campus where the experiment was conducted, so student participants were familiar with the store. A price tag was placed under each item and affixed to the shelves such that participants could not move tags around to cross-examine product information. These tags mimicked the "real" tags found in Staples. Each tag listed the product price and a summary of product features. Unavoidably, some of the summary information on the tags exceeded or was not covered by the RAs; however, the effect of this was negligible, given that the additional information was mostly comprised of the manufacturers' promotional phrases - e.g., "Compact Digital Photo Printer with Dye Diffusion Thermal Transfer Technology" - a language incomprehensible to most participants recruited specifically because of their limited product knowledge. Given that this type of information is provided on price tags in most retail stores, the exclusion of such phrases from the tags themselves would have detracted from the mundane realism of the store setting (Singleton and Straits 1999). Because the benefits of inclusion outweighed the disadvantages of exclusion, I chose to include the phrases on the tags even though they would be missing from the RAs. 102 4.3.8 Participants and Experimental Procedures A total of 75 students at a large North American university participated in the experiment. Twenty five participants were randomly assigned to each of the three treatment groups. The participants' previous knowledge of the product category was carefully controlled by recruiting only those who (1) had not purchased a printer in the last two years, (2) had not previously owned the same printer models, and (3) had little or no expertise with printers (1-4 on a 7-point printer expertise scale ranging from 1 [not at all expert] to 7 [extremely expert]). As in the experiment detailed in Chapter 3, it was important for this experiment to recruit participants with only limited knowledge of the product, because only those without sufficient knowledge of the product, and without clearly established preferences towards the product, would seek advice from RAs (Haubl and Trifts 2000; Swaminathan 2002). Each participant was guaranteed monetary compensation for his/her participation ($20). In order to motivate participants to consider the experiment a serious shopping session, and to increase their involvement, the top 25% performers were offered an additional incentive of $40. Participants were told before the experiment that they would be required to justify their choices and that their performance would be judged based on these justifications. The main criteria for judgment were: (1) how far a participant's choice of printers satisfied his/her family member's and best friend's needs, and (2) whether or not the chosen printers stayed within the price range they had originally set. The experimental session began with the training session as follows (see Figure 4.7). First, each participant watched a tutorial video that described how to use a given RA. Then, a research assistant trained participants to use each of the features of the particular R A in the specific order described in Section 4.3.2, RA Design, above. While each participant was going through the list, a research assistant conducted a clickstream analysis to show which pages participants visited in what order (i.e., monitoring the 103 succession of mouse clicks made by each participant). If the results of the clickstream analysis indicated successful completion by the participant, the training session ended. If a participant failed to complete the task, s/he was asked to perform the task again until s/he had successfully tested all the features. The training session served two purposes. First, because many participants had never used a handheld device before, it was imperative to make sure that each participant had or acquired the necessary skills, such as entering input with a stylus. Second, the session ensured that each participant was fully able to comprehend and utilize the particular decision strategies each R A was designed to support. - INSERT [Figure 4.7] H E R E -Once participants fully understood how to manipulate the R A and the interfaces, they were asked to complete their shopping in two stages (see Appendices A.4 and A . 5 for the detailed experimental tasks). For Stage 1, participants were asked to walk around the store for as long as they wanted and to familiarize themselves with the products without the RA. Once they felt ready, they exited the store and the research assistant provided them with a handheld device on which the R A was running. For Stage 2, participants walked back into the store and shopped for a printer for as long as they wanted. The two-stage shopping process was devised to ensure that the reasonable assumption - i.e., that normal consumers would look around the store and inspect the products available in the store - was realized in the experimental setting. Pilot studies indicated that, without Stage 1, participants were so motivated to use the R A that they did not look around the store, and consequently did not make "in-store" decisions because their shopping experience was entirely dedicated to the use of RAs. In order to minimize any artificiality associated with the two-stage shopping process, the research assistant specifically asked participants to shop as naturally as possible, and to pretend that they had just arrived at the store and were walking around to see what was available on the 104 shelves. Once they were finished with the tasks, participants were asked to complete an online questionnaire. Participants in the control group performed the same tasks except that they were neither trained to use any RAs nor asked to refer to the RAs in the second phase (see Appendix A.5 for the details of the control group task). 4.4 R E S U L T S A N D A N A L Y S E S 4.4.1 Overview SPSS for Windows Version 14.0 was used to conduct analyses of all the dependent variables. Table 4.5 presents the overview of the analyses conducted to compare (1) the control and the R A groups and (2) R A - A L and RA-AT groups. First, the control group was compared with the R A groups in terms of decision accuracy, perceived effort, and intentions to return to the store. Next, the R A - A L group was contrasted against the R A -AT group in terms of number of time that alternatives were viewed, decision accuracy, perceived effort, perceived control, and usage intentions. - INSERT [Table 4.5] H E R E -Table 4.6 shows descriptive statistics for every dependent variable (except decision accuracy measured on an ordinal scale) and the results of reliability testing of the dependent variables using seven-point semantic scales. Cronbach's Alpha was used to assess the reliability of the measures, such as intentions to return to the store, perceived effort, perceived control, and intentions to return to the store. A l l four constructs displayed acceptable levels, .87, .86, .78, and .91, respectively, exceeding the recommended level of .7 (Barclay et al. 1995). ~ INSERT [Table 4.6] H E R E -105 4.4.2 Control vs. R A groups This section presents the results of testing three hypotheses on the differences between the control group and the R A groups in terms of decision accuracy, perceived effort, and intentions to return to the store. Tables 4.7 and 4.8 show a comparison between the number of participants who chose the two most desirable printers and the number of participants who chose other printers across the three groups. In general, the participants in the R A - A L group chose the desirable printers more often than participants in the other two groups. The participants in both the control group and the RA-AT group showed somewhat similar choice patterns. Specifically, for the task "choosing a printer for a family member (i.e., a novice user)," 9 R A - A L users chose the most desirable printer (HP Photosmart 8050), compared to only 3 RA-AT users and 2 control group participants. Also, for the task "selecting a printer for a friend (i.e., an expert user)," 17 R A - A L users chose the most desirable printer (Canon Pixma iP5200), compared to 10 RA-AT users and 12 control group participants. In sum, the differences became salient between the R A - A L group and the rest, contrary to the expectation that the difference between the control group and the two R A groups would be significant. - INSERT [Table 4.7 and 4.8] H E R E -In order to test statistically whether any difference existed between the control and the two R A groups, I conducted chi-square tests. A chi-square test may be used to assess the significance of differences in frequencies on a nominal scale among two independent groups (Siegel and Castellan 1988). It should be noted that the contingency table to be tested was 2 x 2 , that is, two groups and dichotomous choices (i.e., choice of the desirable vis-a-vis the non-desirable printers). A correction for continuity is recommended to prevent overestimation of statistical significance for a 2x2 table with a cell that has an expected count less than 5 (Yates 1934). 106 For the task of choosing a printer for a family member, the minimum expected frequency was 4.7; hence I applied Yates' correction. The result after Yates' correction applied indicates that the frequency with which the most desirable printers were chosen for family members did not differ significantly between the control and the R A groups: X 2 ( l , N = 75) = 1.85, p >.05 (Table 4.9). For the task of choosing a printer for the best friend, the minimum expected count was 12, exceeding the number recommended by Yates (1934). Therefore, a.chi-square test without Yates' correction was conducted; however, the difference was non-significant: X 2 ( l , N = 75) = .24, p >.05. These non-significant results are largely attributable to the small differences between RA-AT and the control group. Consequently, Hypothesis 1, predicting that the R A groups as a whole would make more accurate decisions than the control group, is rejected. - INSERT [Table 4.9] H E R E -Next, the hypotheses regarding perceived effort and intentions to return to the store were tested (Tables 4.10 and 4.11). An A N C O V A was conducted for each variable with the same control variables as in Chapter 3 (see section 3.4.2): consumers' current web usage frequency, previous web experience (Taylor and Todd 1995), gender (Gefen and Straub 1997), tendency to pursue accuracy over effort (Wang and Benbasat 2007), and involvement with printers (Bechwati and Xia 2003). In addition, I included previous experience with using a handheld device to account for the participants' familiarity with handheld devices. R A use significantly affected perceived effort (F(l, 67) = 12.171, p <.01) and intentions to return to the store (F(l , 67) = 13.947, p <.01). These results support hypotheses H2 and H3. - INSERT [Table 4.10 and 4.11] H E R E -107 4.4.3 R A - A L vs. R A - A T Groups This section presents the test results of five hypotheses on the differences between R A - A L and RA-AT user groups, in terms of number of times that alternatives were viewed, accuracy of decisions, perceived effort and perceived control, and intentions to use RAs. The A N C O V A using the same control variables as mentioned above was applied to test the difference in the number of times alternatives were viewed. The results showed that R A - A L users viewed alternatives significantly more often than RA-AT users (F(l , 42) = 11.269,p < .01) (Table 4.12). This results support Hypothesis 4. I measured decision accuracy by observing the share of the two desirable printers among all the chosen printers (Tables 4.9). The minimum expected frequencies were 6.00 (for the task involving a family member) and 9 (for the task involving a friend), larger than the 5 suggested by Yates (1934); I therefore conducted two chi-square tests without applying Yates's corrections (Yates 1934). The R A - A L users chose the desirable printers significantly more often than the RA-AT users for both tasks: X 2 ( l , N - 50) = 3.94, p <.05 (for the task of choosing a printer for a family member); X 2 ( l , N = 50) = 3.94, p <.05 (for the task of choosing a printer for a friend)10. These results support Hypothesis 5, which predicted that R A - A L users would make more accurate decisions than RA-AT users. In addition, I analyzed RA-AT users' clickstreams in order to confirm my expectation that, because of the incompatibility, RA-AT users would revise their preferences repeatedly and would choose the elimination feature, the primary cause of RA-AT users' inaccurate choices. The results show that RA-AT users specified their preferences 3.34 times on average, while R A - A L users revised their comparison set 2.08 1 0 Note that the significance level, .05, was not adjusted. This is because the comparisons between the control group and both RA groups is orthogonal to (uncorrelated with) the comparisons between RA-AT and RA-AL group (Rosenthal and Rosnow 1985). . 108 times. The difference of 1.26 iterations is substantial given that participants performed the iterations with the cumbersome mobile device while standing in the store. Spending a few more minutes can be a daunting job for a participant using the cumbersome mobile device standing in the store, as opposed to using a PC sitting on a chair (Albers and Kim). To further support my anticipation on incompatibility, I analyzed how many RA-AT users utilized the elimination feature and whether the use of the elimination feature accounted for their inaccurate choices. For the task involving a family member, 11 R A -AT users used the elimination feature, and 9 of these 11 chose less desirable printers. In other words, use of the elimination feature accounted for 40% of incorrect choices made by RA-AT users (9 out of 22 RA-AT users who did not choose desirable printers). For the task involving the best friend, 13 RA-AT users used the elimination feature, and 12 of these 13 made less accurate decisions. For this task, the use of the elimination feature accounted for 80% of inaccurate decisions (12/15). These results confirm that, due to the incompatibility, RA-AT users revised their preferences repeatedly and tended to use the elimination feature, and that the use of the elimination feature was one of the main reasons for the difference in decision accuracy between R A - A L and RA-AT users. Three A N C O V A tests were conducted with the same control variables used earlier (Tables 4.13, 4.14, and 4.15). There was no significant difference in terms of perceived effort between R A - A L and RA-AT users (F(l , 42) =.695,p >.05), contrary to Hypothesis 6 which anticipated that R A - A L users would report less effort than RA-AT users. However, R A - A L users reported significantly higher perceived control than RA-AT users (F(l, 42) = 4.702, p <.05), in accord with Hypothesis 7. Lastly, there was no significant difference in terms of intentions to use RAs (F(l, 42) =.025, p >.05), contrary to Hypothesis 8 which predicted that R A - A L users would exhibit higher usage intentions than RA-AT users. - - INSERT [Table 4.13, 4.14, and 4.15] H E R E - -109 4.5 DISCUSSION A N D C O N C L U S I O N S 4.5.1 Summary of Findings Mobile RAs have recently attracted the attention of both academics and practitioners. Mobile RAs, which operate on consumers' personal mobile devices, are designed to provide support for consumers "en route," i.e., when they are already inside a retail store and about to make a product choice. Mobile RAs provide instant access to product information and advice on site, thereby rendering decision making more effective and efficient (Lee and Benbasat 2003; Lee and Benbasat 2004). The purpose of this study is two-fold: The first goal was to investigate whether or not, and in what ways, mobile RAs help consumers to make more accurate decisions, reduce the effort required to make those decisions, and hence increase consumers' intentions to return to a store where they use the RA. The second goal is to examine whether or not compatibility between RA' s guidance and the store's product displays leads consumers to acquire information in a certain way (i.e., to examine alternatives vs. attributes), reduces decision-making effort, increases perceived control, and increases consumers' intentions to use the RAs. With these goals in mind, I conducted a laboratory experiment with one between-subject factor with three levels: control group, R A - A L , and RA-AT. 75 student participants were recruited from a large North American university and randomly assigned to one of the three groups. A simulated store selling 22 printers was created in a laboratory, displaying life-size pictures of printers on shelves. Chi-square tests were used to compare the accuracy of decisions; A N C O V A were used to test the rest of the variables. The results supported some of the expected differences between the control and the R A groups: R A use reduces consumers' effort to make decisions and increases their intentions to return to the store. However, the prediction that the RAs would increase the accuracy of decisions was rejected, due to a large similarity between the RA-AT group and the control group. R A - A L users made more accurate product choices, while RA-AT 110 users showed purchase patterns similar to the control group. The contrast between R A - A L and RA-AT users was found as follows: R A - A L users viewed alternatives significantly more often, reported higher control over manipulating the RAs, and made more accurate decisions than RA-AT users. However, R A - A L users did not exhibit any significant differences from RA-AT users in terms of perceived effort or usage intentions. 4.5.2 Discussion of the Results As expected, mobile RAs reduced consumers' decision-making effort and increased their intentions to revisit the store. These results suggest that the decision-making effort saved by using mobile RAs exceeds the additional effort required to use them simultaneously with shopping in a store. Secondly, the use of mobile RAs increased consumers' intentions to return to the store itself. This is a very interesting result as the mobile RAs were provided by a third party. This indicates that mobile R A users transfer their positive attitudes towards the RAs to the store where they used the RAs even if there is no direct relation between the two. One of the most prominent findings in this study concerns the differences in the accuracy of final decisions. It was expected the R A users, regardless of the type of R A employed, would make more accurate decisions than the control group, given that the RAs provides them with additional product information as well as advanced search and comparison features. In fact, the results show that the type of RAs determines the accuracy of decisions, rather than the use of RAs per se. Whereas R A - A L users chose significantly more desirable products, RA-AT users showed little difference from non-users in terms of accuracy of decisions. This can be attributed to the compatibility between the RAs ' guidance directions and the store's product displays. In sum, what determines the accuracy of purchase decisions is not simply whether or not consumers 111 are provided with mobile RAs, but whether or not the given RAs offer support compatible with the in-store display. Contrary to expectations, there was no difference between RA-AT and R A - A L users in terms of levels of perceived effort and intentions to use RAs. A possible explanation for these non-significant results can be found in the amount of effort R A - A L and RA-AT required of users. That is, R A - A L requires consumers to review products more often than RA-AT does, due to its alternative-driven guidance. It is possible that R A - A L saved users' effort by providing compatible guidance but simultaneously increased their effort to review products. In contrast, RA-AT might have demanded cognitive effort due to its incompatible guidance but helped consumers to reduce the effort to examine products due to its attribute-based approach which is known to save consumers' decision-making effort (Payne et al. 1993). In conclusion, the effort saved by compatible guidance may be exceeded by the effort required to review products. Since the amount of cognitive effort required determines consumers' acceptance of RAs for decision tasks that do not involve products with high emotional content (Payne et al. 1993), it is clear that consumers' intentions to use the RAs did not differ. 4.5.3 Limitations This study has a number of limitations. First, the conceptualization of "desirable" printers versus "less desirable" printers is clearly limited in the sense that, strictly speaking, the desirable printers are not non-dominated products which are not inferior to other alternatives in terms of every attribute. Therefore, it could be argued that the frequency with which desirable printers were chosen may not properly reflect decision accuracy with the same rigour as a scenario involving non-dominated products. I chose to operationalize decision accuracy by observing the frequency with which desirable printers were selected because the use of non-dominated products cannot reveal the difference between the two RAs in terms of decision accuracy. Regardless of 112 which guidance or decision strategies an R A uses, i f the RA' s algorithm works correctly, that R A will select non-dominated products; and therefore differences in decision accuracy cannot be observed. In particular, an R A that employs the E B A strategy will select non-dominated products at all times, since a non-dominated product can never be eliminated because it has higher or equivalent levels for all attributes. Thus, it was not feasible to operationalize decision accuracy by counting the frequency of non-dominated products versus dominated products in this study. Secondly, in any efficient market, non-dominated products do not exist, as conflicts among attributes are inevitable (Haubl and Murray 2003). Therefore, the use of real printers is ecologically valid and increases the realism of the experiment. During the experiment, participants had to deal with complicated, conflicting, and sometimes similar/dissimilar attributes existent in real printers. It is more difficult for a participant to identify desirable printers among less desirable printers than non-dominated products among dominated printers. The former requires the participant to find a printer that satisfies the target person's needs while remaining within his/her budget range, whereas the latter requires him/her to select a printer that has higher attribute levels regardless of these real/realistic concerns. Despite the complexity and difficulty of the tasks, a significant difference in accuracy was found between R A - A L and R A - A T users. This indicates that the effects of compatibility on accuracy are robust. In sum, using desirable vs. less desirable printers instead of non-dominated vs. dominated printers only improved the mundane realism of the experimental task (which is often lacking in such experiments), and did not hamper the validity of the study's measures of decision accuracy. Nevertheless, the use of real printers was inevitably accompanied by the problem of uncontrolled brand image. In other words, participants might already have had strongly positive/negative attitudes towards particular brands, which in turn influenced their 113 selection of printers. However, the results suggest that this factor did not swing the participants' choices, as the desirable printers for the two experimental tasks were of different brands: HP and Canon. R A - A L users chose the HP brand more often for the task involving a family member, and the Canon brand more often for the task involving the best friend. This result indicates that participants did not have strong brand preferences for the two brands of interests (HP and Canon), and that their choices were not determined, by pre-existing preferences towards particular brands but by the particular task. Next, the artificiality of a store setting created in a laboratory should be noted. An artificial store cannot mimic the dynamic real-world purchase situations. Especially, the purchase setting created in the lab might not properly reflect distractions and ambient noises in the real store settings. Nonetheless, distractions implemented in a laboratory can exacerbate the artificiality of the store and furthermore hamper consumers from paying full attention to mobile devices (Pascoe et al. 1998; Pascoe et al. 2000). In addition, I used pictures of printers instead of real printers due to numerous practical reasons including the difficulties in returning printers after the experiment. As I did not have real 22 printers, I could not provide printouts of each printer. However, having the real printers (and printouts) instead of pictures would have only strengthened the results, as participants would have attended to real printers than pictures; hence the effects of compatibility could have been observed more clearly. In addition, many attributes of printers are search attributes (e.g., cartridge cost) consumers can examine sufficiently without direct contact with products (Suh and Lee 2005; Wright and Lynch 1995). Therefore, use of pictures instead of real printers does not invalidate the results. Lastly, this study did not address the issue of unstable wireless Internet connections, which is the primary challenge in developing mobile RAs (Miller et al. 2003). However, as wireless technologies develop rapidly, wireless Internet will become more stable in the 114 near future. 4.5.4 Contributions and Implications Despite its limitations, this study makes a number of contributions to both practice and theory. First of all, this is the first empirical study to investigate the effects of using mobile RAs, equipped with fully functional and interactive preference-elicitation methods, on in-store purchase decisions. The technical feasibility of developing mobile RAs for in-store use had previously been tested in a few Computer Science studies, and Van der Heijden (2002, 2005a, 2005b) had dealt with the in-store use of mobile RAs. However, the Computer Science studies focused on the technical development of RAs rather than the R A s effects on consumers; and Van der Heijden's R A lacked an interactive preference-elicitation method and was therefore limited in terms of reflecting consumers' decision-making constructed "on the fly." In addition, this is the first study that contrasted consumers' perceived effort and intentions to return to the store between mobile R A users and non-users. It was already known that advanced web features, such as RAs, increase consumers' intentions to return to the web store that provides the R A (Jiang and Benbasat forthcoming), but, to date, there had been no studies demonstrating a positive relation between consumers' attitudes towards RAs and a third-party store. Furthermore, this is the first study to look into the issue of compatibility between mobile RAs ' guidance directions and retail stores' product displays. In-store contexts differ from stationary contexts in that in-store R A users must acquire and process information from two different sources: the store and the RA. Consequently, the compatibility between the two becomes important in predicting the positive influence of mobile RAs on consumers' perceptions and acceptance of RAs. Most RAs operated on stationary computers employ an attribute-driven approach, which is thought to decrease consumers' decision-making effort and to fit with consumers' natural approach to solving a shopping task within a stationary computing context. However, this study has shown 115 that an attribute-driven approach is not compatible with the store's product displays and hence is not suitable for in-store purchase contexts. Participants in this study reported that they felt higher control over R A - A L . This study provides the following implications for practice. First, I have shown that consumers' positive attitudes towards RAs are transferred to the store. Based upon this finding, the retailer might want to consider providing wireless Internet inside the store to enable customers to use their own mobile device(s) to review RAs available online, especially i f the retailer lacks the capacity to develop his/her own RA. The recent drop in the cost of wireless Internet makes it more cost effective for retailers to equip their stores with the wireless Internet. Furthermore, the provision of wireless Internet could be particularly beneficial to the retailers who cannot afford to hire and train qualified sales representatives. Mobile RAs can function as surrogate sales representatives by providing product information to consumers. Thus, small-to-medium-sized retail stores may be able cut their sales-representative-related costs by simply providing in-store wireless Internet. Next, mobile R A designers may want to consider the alternative-driven approach, which has been largely neglected in stationary RAs. As the results showed, the alternative-driven approach is more compatible with in-store contexts and may be more effective than attribute-driven RAs. Also, R A - A L does not require as much cost and expertise for designers to develop as RA-AT because R A - A L designers are exempted from the high costs of developing a user-friendly preference-elicitation method (Leavitt 2006). This is because the main role of R A - A L is to support users in making their own side-by-side comparisons. Furthermore, R A - A L generates recommendations by inferring consumers' preferences based upon the products the consumers chose from pair-wise comparisons instead of eliciting consumers' preferences explicitly. In sum, an alternative-driven R A could provide a valuable "alternative" to an attribute-driven R A in that it reduces the cost of development while increasing the quality of service it provides to 116 consumers. 4.5.5 Future Research Future researchers may consider conducting a field experiment to overcome the apparent limitations of investigating in-store decisions in a laboratory. Researchers may want to explore further the effects of compatibility on consumers' decision-making effort in order to resolve the questions left unanswered in this study. More specifically, researchers may want to explicitly compare the effort saved by compatibility and the effort required by using an alternative-driven approach. For example, the positive effects of compatibility on decision-making effort can vary depending on product complexity, such as the number of attributes/alternatives. For complex products that have a number of alternatives and attributes, the effort required to use an alternative-driven approach may exceed the effort saved by compatibility. For less complex products that have a number of attributes but only a few alternatives (e.g., in an oligopolistic market), the effort saved by compatibility may exceed the effort required to use an alternative-driven approach. In addition, future research could examine the moderating effects of product type on the influence of compatibility. In this study, I employed only printers that have many search attributes in order to overcome the apparent limitations of using pictures of products rather than real products. Compatibility, however, may be a more prominent factor in consumers' performance for experience products (which consumers want to sample physically to a greater degree): Consumers will grow more intent on physically trying out experience products than search products; hence they may be more irritated by Fan incompatible R A that does not support their physical product examination. Thus, an investigation of the moderating effects of product type could further develop our understanding of the effects of compatibility. Next, this study investigated the effects of compatibility only in a retail store setting. If both an in-store setting and a stationary-computer-usage setting were 117 simultaneously investigated, the effects of compatibility would be detected more successfully. For instance, whereas RA-AT might be preferred to R A - A L in a stationary computing setting, the difference between RA-AT and R A - A L could decrease in a retail store setting, clearly demonstrating the effects of compatibility. Lastly, research is also needed to investigate whether mobile RAs can support consumers as effectively as sales people. As argued earlier, many consumers tend not to trust sales people; however, they may still prefer interacting with other human beings to relying on technical artefacts, such as RAs. 118 4.6 T A B L E S A N D F I G U R E S Table 4.1 Attributes of Printers Attributes Levels Brand Canon, Epson, Hewlett Packard, Kodak, Lexmark, and Samsung Black and white print speed (in PPM [Page Per Minute]) (1)14 ppm or lower (down to 7 ppm) (2) 15 ppm to 24 ppm (3) 25 ppm or higher (up to 27 ppm) Colour print speed (in PPM [Page Per Minute]) (1)14 ppm or lower (down to 1 ppm) (2) 15 ppm to 24 ppm (3) 25 ppm or higher (up to 33 ppm) Colour resolution (in DPI [Dot Per Inch]) (1) 300 x 300 dpi (2) 4880 x 1200 dpi (3) 5760 x 1440 dpi (4) 9600 x 2400 dpi Print quality (1) Fair (2) Good (3) Very good (4) Excellent Total cartridge costs (1) $81 or more (up to $90) (2) $51 - $80 (3) $21 - $50 (4) $20 or less (down to $18) Maximum printable paper width (1)4 inch (2) 8.5 inch (3) 13 inch Operating systems (1)Mac O S (2) Windows Media card support (1)No (2) Yes Portability (1) No (2) Yes Price (1) $301 or more (up to $400) (2) $201 to 300 (3) $101 to 200 (4) $100CDN or less (down to $39) 119 Table 4.2 Dependent Variables Variables Measures Sources Number of alternatives viewed Number of times a participant clicked on the detailed-information-about-the-product pages Modified from Haubl and Trifts (2000) Intentions to return to the store • Assuming that 1 have access to this store, 1 would like to visit the store next time 1 need to shop for a printer. • 1 would like to visit a store similar to this store next time 1 need to shop for a printer. • 1 would shop at a store that is similar to this store in the future. Coyle and Thomson (2001) Decision Accuracy Share of the two most desirable printers Modified from Haubl and Trifts (2000) Perceived Effort • The task of selecting a printer (using the shopping advisor) took too much time. • Selecting a printer (using the shopping advisor) required too much effort. • The task of choosing a printer (using the shopping advisor) was easy. (R) • The task of selecting a printer (using the shopping advisor) was too complex. Bechwati and Xia (2003) Perceived Control • When specifying my preferences for printers, 1 felt 1 was in control. • 1 think that 1 had a lot of control over the preference-specification process. • The way 1 indicated my preferences for printers made me feel 1 was in control. Bechwati and Xia (2003) Usage Intentions • Assuming 1 have access to the shopping advisor, 1 intend to use it next time 1 consider buying a printer. • Assuming 1 have access to the shopping advisor, 1 predict 1 would use it next time 1 plan to purchase a printer. • Assuming 1 have access to the shopping advisor, 1 plan to use it next time 1 consider buying a printer. Modified from Wang and Benbasat (2007) 120 Table 4.3 Decision Tasks and Desirable Printers Decision Tasks* To choose a birthday present for your family member To choose a wedding gift for your best friend Characteristics of the Person - Occasional home use - Little knowledge and/or experience of colour inkjet printers - Frequent use related to work - Considerable knowledge and experience of colour inkjet printers. Desirable Printer HP Photosmart 8050 Canon iP Pixma 5200 Characteristics of the Printer - Base model - Highest colour print speed. - Second highest levels in terms of cartridge cost and print quality. - All-average levels for other attributes - Advanced model - Highest colour resolution, print quality, and maximum black and white print speed - Second highest levels in terms of cartridge cost and price. *The order of the decision tasks was counterbalanced. 121 Table 4.4 Attribute Levels of 22 Printers Portable Media Maximum Maximum Maximum Maximum Print Cartridge Sum of Price ($) Card Paper Color Print Color Black & White Quality Cost Attribute Slot Width Speed Resolution Print Speed Levels H P Deskjet 9800 1 1 3 2 2 3 3 2 17 399 .95 Canon Pixma iP6600D 1 2 2 2 4 2 2 1 16 269 .92 Epson Picture Mate 2 1 1 3 0 2 3 14 2 4 9 . 9 3 Epson Stylus Photo R340 1 2 2 2 3 2 2 1 15 249 .92 H P Photosmart 8250 1 2 2 3 2 3 2 2 17 249 .92 H P OfficeJet Pro K550 1 1 2 3 2 3 2 2 16 2 2 9 . 9 5 Canon Pixma iP5200 1 1 ^ 2 2 4 3 4 3 20 199.96 Canon Selphy DS810 1 2 1 1 2 0 1 4 12 179.92 Canon Pixma iP4200 1 1 2 2 4 3 1 1 15 149.36 H P DeskJet 6940 1 1 2 3 2 3 2 1 15 144.86 Canon Selphy CP510 1 1 1 1 0 2 4 12 129.96 H P DeskJet 5940 1 1 2 2 2 3 2 2 15 129.42 Canon Pixma JP6220D 1 1 2 1 2 0 3 3 13 127.00 HP Photosmart 8050 1 2 3 2 3 '• 3 - - ,; 125.46 Lexmark P315 1 2 1 1 2 0 3 3 13 99 .93 Epson Stylus C88 1 1 2 1 3 2 1 1 12 99 .92 H P Photosmart 335 2 1 1 2 0 3 3 14 99 .92 Canon Pixma iP6210D 1 1 2 1 2 0 3 3 13 97 HP DeskJet 5440 1 1 2 1 2 1 3 3 14 89.91 Canon Pixma iP1600 1 1 2 2 2 2 2 2 14 79 .73 Samsung S P P 2020 Photo Printer 2 1 1 1 1 0 3 3 12 69 .98 Lexmark Z517 1 1 2 1 2 1 1 1 12 39 .98 Table 4.5 Overview of Hypotheses Testing Comparisons Control vs. RA Groups RA-AL vs. RA-AT Dependent Variables Intentions to Return to the Store Tested -Decision Accuracy Tested Tested Perceived Effort Tested Tested Number of Times Alternatives were Viewed - Tested Perceived Control - Tested Intentions to Use the RA - Tested Table 4.6 Descriptive Statistics Cronbach's Alpha Means and Standard Deviations Com parisons Control RAs RA-AL RA-AT Intentions to return to the store* .87 4.65 (1.04) 5.51 (.80) N/A Perceived Effort* .86 3.70 (.62) 3.18 (.55) 3.24 (.65) 3.13 (.44) Number of alternatives viewed N/A N/A 24.36 (21.08) 10.80 (9.47) Perceived Control* .78 5.89 (.76) 5.32 (1.02) Intentions to use the RA* .91 5.93 (.60) 5.97 (.89) * indicates the variables measured on seven-point semantic scales. 123 Table 4.7 Printers Chosen for a Family Member Printer models Price ($) Groups Total RA-AL RA-AT Control Desirable printer HP Photosmart 8050 125.46 9 3 2 14 Less Lexmark Z517 39.98 1 2 3 desirable Samsung S P P 2020 Photo 69.98 1 1 2 printers Canon Pixma iP1600 79.73 2 1 2 5 HP Deskjet 5440 89.91 3 2 1 6 Canon Pixma iP6210D 97.00 4 6 10 Epson Stylus C88 99.92 5 5 HP Photosmart 335 99.92 5 3 8 Canon Pixma iP6220D 127.00 1 2 1 4 HP Deskjet 5940 129.42 1 3 4 HP Deskjet 6940 144.86 1 1 2 Canon Pixma iP4200 149.36 1 1 Canon Selphy DS810 179.92 1 1 Canon Pixma iP5200 199.96 3 2 1 6 HP OfficeJet Pro K550 229.95 1 1 Epson Stylus Photo R340 249.92 1 1 HP Photosmart 8250 249.92 1 1 2 Total 25 25 25 75 Table 4.8 Printers Chosen for a Friend Printer models Price ($) Groups Total RA-AL RA-AT Control Desirable printer Canon Pixma iP5200 , ' 199-96 ' - 17 10 : 12 ' 39 Less desirable printers HP Deskjet 5440 89.91 1 1 Canon Pixma iP6210D 97.00 1 1 Epson Stylus C88 99.92 1 1 HP Photosmart 335 99.92 1 1 1 3 Lexmark P315 99.93 1 1 HP Photosmart 8050 125.46 2 3 3 8 HP Deskjet 6940 144.86 2 1 3 HP OfficeJet Pro K550 229.95 1 1 2 Epson Stylus Photo R340 249.92 1 2 3 6 HP Photosmart 8250 249.92 3 2 5 Epson Picturemate 249.93 1 1 Canon Pixma iP6600D 269.92 2 2 HP Deskjet 9800 399.95 2 2 Total 25 25 25 75 124 Table 4.9 Chi-Square Test Results Control vs. RA Groups RA-AL vs. RA-AT Minimum expected count Chi-square value P Minimum expected count Chi-square value P Printer-choice for a family member 4.70 1.85 .09 6.00 3.94 .04* Printer-choice for a friend 12.00 .24 .62 9.00 3.94 .04* * indicates p < .05 Table 4.10 ANCOVA Results: Perceived Effort Source Sum of Squares df Mean Square F Sig. Corrected Model 5.368 ( a ) 7 .767 2.184 .047 Intercept 4.647 1 4.647 13.233 .001 Tendency to pursue accuracy over effort .001 1 .001 .003 .953 Previous experience with handheld devices .065 1 .065 .186 .667 Previous web experience .236 1 .236 .671 .415 Gender .586 1 .586 1.670 .201 Current web usage frequency .043 1 .043 .121 .729 Product Involvement .168 1 .168 .479 .491 RA Use (Use vs. Non Use) ;.. .4 .274 1 4.274 12.171 .001** Error 23.528 67 .351 Total 873.938 75 Corrected Total 28.897 74 (a) R 2 = .186 (Adjusted R 2 = .101); * indicates p < .05; ** indicates p < .01 125 Table 4.11 ANCOVA Results: Intentions to Return to the Store Source Sum of Squares df Mean Square F Sig. Corrected Model 17.486 ( a ) 7 2.498 3.199 .006 Intercept 6.315 1 6.315 8.085 .006 Tendency to pursue accuracy over effort .825 1 .825 1.057 .308 Previous experience with handheld devices 1.297 1 1.297 1.661 .202 Previous web experience .589 1 .589 .754 .388 Gender .905 1 .905 1.159 .286 Current web usage frequency 1.642 1 1.642 2.102 .152 Product Involvement .988 1 .988 1.265 .265 RA Use (Use vs. Non Use) 1 "i 10.892 1 10.892" fcb^3l3j947'' .. .000** Error 52.327 67 .781 Total 2118.667 75 Corrected Total 69.813 74 (a) R 2 = .250 (Adjusted R 2 = .172); * indicates p < .05; ** indicates p < .01 Table 4.12 ANCOVA Results: Number of Times Alternatives Were Viewed Source Sum of Squares df Mean Square F Sig. Corrected Model 5214.322(a) 7 744.903 3.158 .009 Intercept 98.970 1 98.970 .420 .521 Tendency to pursue accuracy over effort 1155.044 1 1155.044 4.897 .032 Previous experience with handheld devices 1169.417 1 1169.417 4.958 .031 Previous web experience 940.304 1 940.304 3.987 .052 Gender 247.536 1 247.536 1.050 .311 Current web usage frequency 1201.272 1 1201.272 5.093 .029 Product Involvement 22.961 1 22.961 .097 .757 RA Type.(RA-AL vs ; RA-AT) 2657.855 * ;2657:855 .002** Error 9905.858 42 235.854 Total 30573.000 50 Corrected Total 15120.180 49 (a) R Squared = .345 (Adjusted R Squared = .236); * indicates p < .05; ** indicates p < .01 126 Table 4.13 ANCOVA Results: Perceived Effort Source Sum of Squares df Mean Square F Sig. Corrected Model 1.704(a) 7 .243 .763 .621 Intercept 4.696 1 4.696 14.721 .000 Tendency to pursue accuracy over effort .011 1 .011 .035 .852 Previous experience with handheld devices .109 1 .109 .343 .561 Previous web experience .462 1 .462 1.449 .235 Gender 1.052 1 1.052 3.299 .076 Current web usage frequency .016 1 .016 .049 .826 Product Involvement .011 1 .011 .035 .851 . RA Type (RA-AL vs: RA-AT) .222 .222 .695 .409 Error 13.398 42 .319 Total 522.313 50 Corrected Total 15.101 49 (a) R 2 = .113 (Adjusted R 2 = -.035); * indicates p < .05; ** indicates p < .01 Table 4.14 ANCOVA Results: Perceived Control Source Sum of Squares df Mean Square F Sig. Corrected Model 7.878(a) 7 1.125 1.319 .265 Intercept 5.337 1 5.337 6.256 .016 Tendency to pursue accuracy over effort .161 1 .161 .189 .666 Previous experience with handheld devices .763 1 .763 .894 .350 Previous web experience -.997 1 .997 1.169 .286 Gender .354 1 .354 .415 .523 Current web usage frequency 2.110 1 2.110 2.473 .123 Product Involvement .663 1 .663 .777 .383 RA Type (RA-AL vs. RA-AT) 4.011 1 4.011 4.702 .036* Error 35.831 42 .853 Total 1615.444 50 Corrected Total 43.709 49 (a) R 2 = .180 (Adjusted R 2 = .044); * indicates p < .05; ** indicates p < .01 127 Table 4.15 ANCOVA Results: Usage Intentions Source Type III Sum of Squares df Mean Square F Sig. Corrected Model 4.655 ( a ) 7 .665 1.208 .320 Intercept 3.844 1 3.844 6.982 .012 Tendency to pursue accuracy over effort .221 1 .221 .402 .529 Previous experience with handheld devices 2.058 1 2.058 3.738 .060 Previous web experience .752 1 .752 1.366 .249 Gender .346 1 .346 .629 .432 Current web usage frequency .023 1 .023 .041 .840 Product Involvement .774 1 .774 1.406 .242 RA Type (RA-AL vs. RA-AT) .014 1 .014 .025 .875 Error 23.125 42 .551 Total 1799.889 50 Corrected Total 27.780 49 a R 2 = .168 (Adjusted R 2 = .029); * indicates p < .05; ** indicates p < .01 128 Figure 4.1 Diagram of R A - A L (1) Introduction Page (2) Product-list page (5) Comparison Page - ADDIF, MCD, and SAT strategies 1. The user clicks on the printers for which s/he wants to view details 3. The user indicates whether s/he wants to compare the printer against other printers in which s/he is also interested. 5. RA-AL recommends six printers that are similar to the remaining few obtained from the pair-wise comparisons. (3) Detailed-lnformation-About-Product Page 2. The user can check the definitions of the attributes contained in this page. (4) Information-about-Attribute page (Same page, but the user can return either to the product-list page or the recommendation page as shown.) 7. The user can compare the recommended printer against the few remaining printers. (6) Recommendation Page • EQW strategy 6. The user can check the details of the recommended printer. (7) Detailed-lnformation-About-Product Page 4. The user compares printers pair-wise until a few printers remain. 8. T h e user m a k e s the final choice. to Figure 4.2 Screenshots of RA-AL IPAQ elTfr P r i n t e r A d v i s o r . c o m 0 for your personalized. unbiased shopping advice you A N S W E R a few simple questions i WA w E N T E R S I T E we GUIDE you to the right printers m .PAQ Pocket PC PHnterAdvisor.com Click each printer to see details wL Check printer boxes to compare 1 o f 2 H C o m p a r e s • Lexmark Z517 • 39.98 • • Samsung S P P 2020 Plum mil • • C u i o n Pixma I P I I > O O • 79.73 • • H P Deskjet 5140 •89.91 • • Canon PixiiialPl>210D •97.00 • • Epson Stylus C88 $99.92 • • H P PhotoSmart 335 •99.92 • • L e x m a r k 1*315 •99,93 • • H P PhotoSmart 8050 • 125.46 • • Canon Petma ,pt>220D • 127.00 • • H P Deskjet 5040 * 129.42 • |i»iy;w^^i^B^Baaaaaaaaaaaaj^^|^|^| m GEO <& iPAQ Lexmark Z517 $39.98 Add to Cart Pocket PC Proa 1 low cost, low noise-level. C o n s Expensive ink lepl.icement cost, pi int quality only f .111. Summary Economical ptintei Th.it can get Ks job done. Bew.11 • of the ink replacement cost. m Introduction Page Product4Jst Page Detailed-lnformation-about-Product Page What is Operating System ? > ft is important that your printer it compatible with your computer's Operating System (OS}. Here we list a l l OS each printer is compatible with. Note that printers compatible with WtnXP are also compatible with older Windows versions, including Win98/20D0/ME. C l o s e S IPAQ Pocket PC 1012 0 Remove Remove Model Lexmark Z517 Lexmark P315 • 39.98 1.99.9! 4200x1200 l| 4800x1200 II 2400x1200 1 MA 1 7 II 12 1 w II IIA 1 r.iir Very Good •84.92 MSM 8.5 in. 4 in. Mac. Windows Mac, Windows Mo Yee Mo Ho More like this More like this Jl IAdd to • Add to • Comparison Comparison Price •99.92 •89.91 Max Ras 5700x1440 | | 4800x1200 | | (dpi) 5700x1440 | 1200x1200 | Max Spd 12 II s II (ppm) 22 • 7.4 | Print Quality f.in Very Good •80.84 •f 44.89 8.5 in. 8.5 in. am Mac. Windows Mac. Windows Ho Ho CaidSlot Mo Me Portable \n View Cart Close M m CS5 Information-about-Attribute Page Comparison Page Recommendation Page 130 Figure 4.3 Diagram of RA-AT (1) Introduction Page (2) Attribute-List Page 1. The user clicks on the attributes for which s/he wants to view details. (3) Information-about-Attribute Page 3. RA-AL recommends six printers that meet user-specified preferences. (4) Preference-Specification Page 5. The user can modify his/her specified preferences. (5) Recommendation Page - Simplified WAD and EBA strategy 4. The user can check the details of the recommended Drinter. (6) Detailed-lnformation-About-Product Page 2. The user indicates his/her attribute preferences and the importance of these attributes in his/her purchase. t 6. The user makes the final choice. Figure 4 .4 Screenshots of RA-AT iPAQ Pocket PC PrrnterAdvisor.com ™ for your personal tied, unbiased shopping advice you A N S W E R a tew simple questions ENTER SITE G U I D E you to the right printer* ES iPAQ Pocket PC PrinterAdvisor.com C h o o s e a feature to specify CL your preferences View Results '•:•} • Price O • Brand O • B&W Print Speed O • Color Print Speed O • Color Resolution O • Print Quality O • Total Cartridge Cost ii • Man Paper Width 0 • Operating System il • Media Card Slot O • Portable ii 11 IPAQ Pocket PC PrinterAdvisor.com C h o o s e a feature to specify CL your preferences View Results 0 I Pr ice O I B r a n d 0 I B & W Print Speed a What is Price? > Approximate retail price, based on information from manufacturers and from national price surveys. Closeg • Mas Paper Width O • Operating System il • Media C a r d Slot 0 • Portable O Introduction Page Attribute-List Page Information-about-Attribute Page iPAQ Pocket PC PrinterAdvisor.com Choose a feature to specify CL your preferences View Resul ts H > Price O I Bra nd O I B & W Print Soeed Q > Price should be between... - s e l ec t -> C o m p a r e d to other features, price is... • Operat ing S y s t e m O • Media C a r d Slot © • Portable O iPAQ Pocket PC •cr . iM.itt h 98% match H M i M +127.00 9600x?4O« 11 600x600 | 4800x1700 t1 4800x1?00 I 24 | | 30 1 1 II D.it.. II A 1 Excellent Veiy Good f J4.9C IJ4 .M 1.5 iii. 8.5 in. Mac. Windows M.n . WllKlOAS PoctetPC Canon Pixm-i iP5200 $199.96 A d d to Cart D i m e n s i o n s I W x H x D l © W e i g h t a M a x S p e e d I p p m l © M a x R e s o l u t i o n (dpi) © Print Qual i ty © Total Cartr idge C o s t © C a r t r i d g e T y p e s ii 17 .Sx6 . )x12 .2 in. 16.1 Ills ?4 II i )0 | •'••«<> • ' I'M) | | I | Excellent • 14.96 Preference-Specification Page Recommendation Page Detailed-lnformation-about-Product Page 132 Figure 4.5 Experimental Store Pi: picture of printers ,: 1-22 Printers are randomly displayed on three shelves Notepad where participants indicated their final choices P8 P7 P6 P5 P4 P3 P2 P1 Door Windows P 22 P 21 P 20 P 19 -P9-P10 P11 P12 P13 P14 P15 P16 P17 P 18 Price tag 133 Figure 4.6 Picture and Price Tag of a Printer (Samsung 2020) Samsung SPP 2020 Photo Printer ^ Compact Digital Photo Printer with Dye Diffusion Thermal Transfer Technology > Up to 4800 x 1200 dpi resolution > Memory: 1MB Flash, 16MB D R A M ^ Long lasting edge-to-edge borderless photos (4" x 6",5"x7" & 8-1/2" x 11") > Connectivity: USB 2.0, PjcXBndge, Wireless printing capable from select devices > Windows 98/00/Me/XP, Mac OS 10.1.5/10.2/10.3 y 1-year or 2000 print Warranty (whichever comes first.)| Our Price: $69.98 134 Figure 4.7 Experimental Procedures 1. Background Questionnaire Par t ic ipants fill out the backg round n i iest innnai r f i . 2. Tutorial Par t ic ipants wa tch the tutorial v ideo that exp la ins how to u s e the shopp ing adv isor . 3. Shopping Task 1 A . P roduc t inspect ion W I T H O U T the R A B. C h o i c e W I T H a he lp of the R A 4. Shopping Task 2 A . P roduc t inspect ion W I T H O U T the R A B. C h o i c e W I T H a help of the R A Par t ic ipants per form 7WO two -phased shopp ing t asks , e a c h of wh i ch invo lves the se lec t ion of a printer. Fo r e a c h task, a part ic ipant w a s a s k e d to c h o o s e one printer f rom a m o n g those ava i lab le ins ide the shopp ing s tore. 5. Post-Shopping Questionnaire A n s w e r ques t ions b a s e d on the part ic ipant 's shopp ing expe r i ence . 135 CHAPTER 5: CONCLUSIONS AND FUTURE RESEARCH 5.1 C O N C L U S I O N S RAs have the potential to support consumers in reducing information overload and task complexity, and facilitate accurate product choices (Maes et al. 1999; O'Keefe and McEachern 1998). A rich body of research has attended RAs since their introduction by GroupLens project by the University of Minnesota (www.cs.umn.edu/research/GroupLens/index.html) in early 90. Most research on RAs focused on underlying calculations and algorithms that generate recommendations. However, consumers' perceptions and acceptance of RAs are influenced by many other factors, such as product- and context-related issues (Todd and Benbasat 1992; Xiao and Benbasat 2007). I focused on product type, particularly, products with emotional contents (in Study 1) and in-store context (in Study 2) among many understudied factors. Product type is an important influence on consumers' decision making processes and outcomes (Levin et al. 2003; McCabe and Nowlis 2001). In-store context becomes relevant and important with the advent of mobile RAs, which assist in-store consumers' in reaching a product choice at the point of purchase (Van der Heijden 2005). Study 1 concerned the effects of attribute conflicts (the prominent characteristics of products with high emotional content) on consumers' perceived control, effort, recommendation quality, and intentions to use the RAs. The W A D and E B A strategies are on the two extremes in presenting the conflicts explicitly or implicitly. Hence, two RAs that employ W A D and E B A individually were implemented and contrasted. A controlled laboratory experiment was conducted with 100 participants who were recruited from a large North American university. The A N C O V A and the separate t-tests results show that R A - W A D highlights attribute conflicts, thereby negatively influencing 136 consumers' perceived effort, quality of recommendation, and intentions to use the RAs. In addition, task emotionality moderates such relationship. Consumers are more motivated to cope with high-emotion tasks, which are associated with severe negative consequences, than with low-emotion tasks, which they connect with little harm to themselves. As a result, the negative effects of attribute conflicts on consumers' perceptions and usage intentions are more pronounced for high-emotion tasks, while these effects are limited for low-emotion tasks. There was no significant main effects of R A type or interaction effects between R A type and task emotionality on perceived effort, the latter, however, as PLS results showed, was not influential on consumers' strategy choice for emotion-laden products (Luce 1998; Luce et al. 2001). Study 2 dealt with in-store contexts where compatibility between the RA ' s guidance directions and the store's product displays becomes an important predictor of consumers' decision performance (Slovic 1972; Slovic et al. 1990). Users of R A - A L that is more compatible with the store's display made more accurate decisions and perceived higher control than those of R A - A T . However, there was no significant difference in terms of perceived effort and intentions to use the RAs. As effort is a determining factor in general decisions involving non-emotional products, in contrast to the case of emotion-laden products in Study 1, consumers did not show significant differences in usage intentions. In addition, the comparisons between the mobile R A groups and the control group demonstrate that R A use reduces consumers' decision-making effort and increases their intentions to return to the store. However, the prediction that the RAs as a whole would increase the accuracy of decisions was rejected, due to a large similarity between the R A - A T group and the control group. Study 1 and 2 together are in accord with the constructive decision-making theory that asserts that consumers' preferences for decision strategies are constructed on the fly contingent upon many contextual factors (Payne et al. 1993). For decisions involving 137 emotion-laden products, consumers value the goal of minimizing confrontations with attribute conflicts even at the expense of the goal of minimizing effort; therefore, their perceived effort did not influence their intentions to use the RAs. On the other hand, consumers value the minimization of effort for decisions involving non emotion-laden products; hence, in Study 2, their perceived effort was the primary influence on their intentions to use the RAs. 5.2 L I M I T A T I O N S Besides the limitations stated in the Chapters 3 and 4 previously, there are three limitations common in Studies 1 and 2. The three limitations are associated with mainly the use of controlled laboratory experiments. First, Studies 1 and 2 employed controlled laboratory experiments which did not involve participants' actual spending of their own money. In other words, participants were told to imagine that they were making real choices in either the website (Study 1) or the store (Study 2). These were hypothetical purchases not real purchases where they had to pay out of their own pockets. Such hypothetical purchases do not mimic the intensity of real-world purchases involving money-spending and risks of inaccurate choices. Clearly these studies are limited in this sense. However, the two studies required close control of other factors that occur in real-world purchases, such as source credibility of websites (Kim and Benbasat 2006), influence of salespeople (Mallalieu 2006) and store environment, such as background music (Morin, Dube and Chebat 2007). Therefore, the choice of controlled laboratory experiments was selected despite the clear limitations of the hypothetical purchases. Second, I needed to ask the participants to justify their choices upon completion of their tasks in Studies 1 and 2, in order to overcome the clear limitations in the hypothetical purchases. Participants in a hypothetical purchase situation could be less 138 motivated in making accurate purchase choices than real-life choices. In order to increase their motivations in the tasks, I requested participants to provide their written justifications about their final choices and told them that top 25% of performers would receive additional incentives of $40, twice more amount than the guaranteed incentive of $20. The evaluation criteria were explicitly mentioned to the participants: whether the justification was convincing and logical in Study 1; whether the choices satisfied both the needs of the persons they had made choices for and whether the choice stayed within their budget ranges. Some might argue that pressure to provide justifications could have affected their choice patterns and resultant perceptions of the RAs. However, anyone who do not live entirely alone has to justify his/her choices (Huber and Seiser 2001). Especially, the experimental tasks in both studies involved choice of products not for participants themselves but participants' family members or friends. Therefore, i f they had made real purchases, they would have needed to explain their choices to the persons they made choices for. Specifically, it is not hard to expect one to explain explicitly why s/he chose the car for his/her family member. Likewise, when giving presents, one is likely to explain why s/he chose the particular present for them. Moreover, Huber and Seiser (2001) has shown empirically that pressure for justifications does not change consumers' choice of decision strategies or alter their decision processes as opposed to when there is no need for a justification. The pressure to justification only leads to a distinct increase in the amount of utilized information and to a more elaborate choice process (Huber and Seiser 2001). This is precisely what I attempted to achieve by requiring the participants to justify their choices: I aimed to motivate participants' to use the given RAs carefully and conscientiously so that they could evaluate advantages and disadvantages of the RAs. Lastly, the experimental tasks of Studies 1 and 2 required participants to reach a product choice, rather than to simply gather information about products. RAs not only 139 assist consumers in immediate purchases on the spot but also in collecting information for later purchases either in the RA's websites or offline stores affiliated with the R A s websites (Leavitt 2006). One of the RA's prominent roles is to provide information to mere web browsers in an attempt to transform them into potential buyers (Leavitt 2006). Such role was not covered by either of Study 1 or 2. Nonetheless, the focus of this dissertation was on role of RAs in assisting consumers' immediate purchase decisions at the point of purchase, rather than in providing information for future purchases. Therefore, such limitation was unavoidable. 5.3 CONTRIBUTIONS In addition to contributions made by individual studies, the two studies make two major contributions collectively: (1) the extension of R A literature into the moderating effects of product type and in-store context and (2) the provisions of R A design guidelines for practitioners. This dissertation illustrated how product type and in-store setting alter consumers' perceptions and usage intentions of RAs. This is in line with the constructive decision-. making theory which asserts consumers' adaptive choice of strategies affected by tasks and contexts. More specifically, consumers' effort-accuracy goals were known to be the determinants of their acceptance of RAs; this dissertation, in contrast, showed that consumers' goal of avoiding attribute conflicts dominated their effort-accuracy goal. Also, consumers were believed to prefer an attribute-driven approach over an alternative-driven approach; in fact, consumers performed better with an alternative-driven approach than with an attribute-driven approach when making a purchase decision in a store. This dissertation clearly demonstrates that consumers' perceptions and acceptance of RAs must be understood in conjunction with many factors associated with their decisions, particularly, product type and in-store context. Therefore, this dissertation extends R A 140 literature and broadens the understanding of consumers' perceptions and acceptance of RAs Second, this dissertation provides design guidelines to practitioners. Study 1 demonstrated that consumers' perceptions of recommendations do not reflect the objective accuracy of recommendations for emotion-laden products. Such an overestimation of accuracy resulted from using the R A - E B A has been reported consistently by many empirical studies (Kotteman et al. 1994, Aloysius et al. 2006, and Todd and Benbasat 1999). This consistent result suggests that a further refinement of RA' s algorithms may not change the consumer's fundamental biases (Griffin and Tversky 2002). This suggests that the designer's primary goal should be not just to develop the RA' s algorithms that increase objective accuracy but to devise the R A that is perceived to be accurate. Therefore, R A - E B A may be a good option for products with high emotional content. In a similar vein, the second study shows that designers may want to take into account how mobile RAs complement the context, such as in-store setting, where the mobile RAs are used, rather than focusing on the developments of accurate algorithms. The task of examining actual products in a retail store may not be as difficult as, or more pleasing than, the task of reviewing products online. Therefore, in-store consumers may not appreciate the RA' s role of recommending products because they can (or sometimes even prefer) select(ing) several products that particularly interested them on their own. What lacks in the in-store context is rather the resources for consumers to compare the selected products (Youll et al. 2001), precisely due to the difficulty of integrating inconsistent information presented on the price tags in the store (Russo 1977; Russo et al. 1986). Therefore, in-store consumers would appreciate the RA' s assistance in comparing products than in providing recommendations. In sum, designers for mobile RAs should consider what assistance consumers need the most in the context where the consumers 141 use the mobile RAs and develop the RAs that respond to the consumers' needs. 5.4 FUTURE RESEARCH As described in Sections 3.5 and 4.5, there are many future research areas which are also important and promising. First, researchers may investigate how the preference-elicitation method of RA-WAD could be designed to alleviate its conflict-confronting nature, so as to encourage consumers to use the desirable and compensatory strategy. Researchers may also take a closer look at perceived control, given that the results of this study suggest that it is not actual control but an illusion of control that determines users' confidence in dealing with attribute conflicts. Researchers may want to explore further the effects of compatibility on consumers' decision-making effort. Specifically, researchers may want to explicitly compare the effort saved by compatibility and the effort required by using an alternative-driven approach. For example, the positive effects of compatibility on decision-making effort can vary depending on product complexity, such as the number of attributes/alternatives. In addition, future research could examine the moderating effects of product type on the influence of compatibility. Compatibility may be a more prominent factor in consumer performance for experience products, as consumers want to sample experience products physically to a greater degree. Next, a future study can compare the effects of compatibility across in-store setting and a stationary-computer-usage setting. Lastly, research is also needed to investigate whether mobile RAs can support consumers as effectively as sales people. 142 R E F E R E N C E S Agarwal, R., and Karahanna, E. "Time Flies When You're Having Fun: Cognitive Absorption and Beliefs about Information Technology Usage," Management Information Systems Quarterly (24:4), 2000, pp 665-694. Albers, M.J. , and Kim, L . "User Web Browsing Characteristics Using Palm Handhelds for Information Retrieval," the 18th annual ACM International Conference on Computer Documentation: Technology & Teamwork, IEEE Educational Activities Department, Cambridge, Massachusetts, 2000, p. 125. Aloysius, J.A., Davis, F.D., Wilson, D.D., Taylor, R.A., and Kottemann, J.E. "User Acceptance of Multicriteria Decision Support Systems: The Impact of Preference Elicitation Techniques," European Journal of Operational Research (169:1), 2006, pp 273-285. Armstrong, G., and Kotler, P. Marketing: An Introduction, CJ ed.) Prentice Hall, 2004, p. 720. Averill, J.R. "Personal Control over Aversive Stimuli and Its Relationship to Stress," Psychological Bulletin (80:4), 1973, pp 286-303. Barclay, D., Thompson, R., and Higgins, C. "The Partial Least Squares (PLS) Approach to Causal Modeling: Personal Computer Adoption and Use as an Illustration," Technology Studies (2:2), 1995, pp 285-309. Baron, R .M. , and Kenny, D.A. "The Moderator-Mediator Variable Distinction in Social Psychological Research: Conceptual, Strategic, and Statistical Considerations," Journal of Personality and Social Psychology (51:6), 1986, pp 1173-1182. Bechwati, N .N. , and Xia , L . "Do Computers Sweat? The Impact of Perceived Effort of Online Decision Aids on Consumers' Satisfaction with the Decision Process," Journal of Consumer Psychology (13:1&2), 2003, pp 139-148. Bettman, J.R. An Information Processing Theory of Consumer Behavior, Addison Wesley, Reading, M A , 1979. Bettman, J.R., Johnson, E.J., Luce, M.F., and Payne, J.W. "Correlation, Conflict, and Choice," Journal of Experimental Psychology: Learning, Memory, and Cognition (19:4), 1993, pp 931-951. Bettman, J.R., Johnson, E.J., and Payne, J.W. "A Componential Analysis of Cognitive Effort in Choice," Organizational Behavior and Human Decision Processes (45:1), 1990, pp 111-139. Bettman, J.R., Luce, M.F., and Payne, J.W. "Constructive Consumer Choice Processes," Journal of Consumer Research (25:4), 1998. Campbell, M.C. , and Kirmani, A . "Consumers' Use of Persuasion Knowledge: The Effects of Accessibility and Cognitive Capacity on Perceptions of an Influence Agent," Journal of Consumer Research (27:1), 2000, pp 69-83. Card, S., Moran, T.P., and Newell, A. The Psychology of Human-Computer Interaction 143 Lawrence Erlbaum Associates, Hillsdale (NJ), 1983. Chernev, A . "Goal-Attribute Compatibility in Consumer Choice," Journal of Consumer Psychology (14:1&2), 2004, pp 141-150. Chin, W.W., Newsted, P.R., and Hoyle, R .H. "Structural Equation Modeling Analysis with Small Samples Using Partial Least Squares," in: Statistical Strategies for Small Sample Research, Anonymous (ed.), Sage publications, Inc., Thousand Oaks, California, 1999, pp. 307-341. Cosley, D., Lam, S.K., Albert, I., Konstan, J.A., and Riedl, J. "Is Seeing Believing? How Recommender Systems Influence Users' Opinions," in: CHI 2003 Conference on Human Factors in Computing Systems, Fort Lauderdale, FL. , 2003. Coyle, J.R., and Thomson, E. "The Effects of Progressive Levels of Interactivity and Vividness in Web Marketing Sites," Journal of Advertising (30:3), 2001, pp 65-78. Dasgupta, S., and Chanin, M . "Knowledge Quality In Knowledge-Based Systems: An Investigation In A Group Decision Making Environment," Americas Conference on Information Systems, Long Beach, CA, 2000. . Davis, F., and Kottemann, J.E. "User Perceptions of Decision Support Effectiveness: Two Production Planning Experiments," Decision Sciences (25:1), 1994, p 57. Delquie, P. "Optimal Conflict in Preference Assessment," Management Science (49:1), 2003, pp 102-115. Dickinson, S., and Barker, A. "Evaluations of Branding Alliances between Non-Profit and Commercial Brand Partners: The Transfer of Affect," International Journal of Nonprofit and Voluntary Sector Marketing (12), 2007, pp 75-89. Diehl, K. "Tempted by the Cheap and Easy: Searching Too Much in Ordered Environments," in: Unpublished Working Paper, Marketing Department, Moore School of Business, University of South Carolina, 2003. Drolet, A. , and Luce, M.F. "The Rational-izing Effects of Cognitive Load on Emotion-Based Tradeoff Avoidance," Journal of Consumer Research (31:1), 2004, pp 63-77. Einhorn, H. , and Hogarth, R. "Confidence in Judgment: Persistence of the Illusion of Validity," Psychological Review (85), 1978, pp 395-416. Fasolo, B. , McClelland, G.H., Lange, K . A . , Haugtvedt, CP. , Machleit, K . A . , and Yalch, R.F. "The Effects of Site Design and Interattribute Correlations on Interactive Web-Based Decisions," in: Online Consumer Psychology: Understanding and Influencing Consumer Behavior in the Virtual World, C P . Haugtvedt, K . A . Machleit and R.F. Yalch (eds.), Lawrence Erlbaum Associates, Inc., New Jersey, 2005, pp. 325-342. Fitts, P.M., and Seeger, C M . "S-R Compatibility: Spatial Characteristics of Stimulus and Response Codes," Journal of Experimental Psychology (46:3), 1953, pp 199-210. Frese, M . , Ulich, E., and Dzida, W. "A Theory of Control and Complexity: Implications for Software Design and Integration of Computer Systems into the Work Place," in: Psychological Issues of Human-Computer Interaction in the Work place, M . Frese (ed.), Elsevier Science, Amsterdam ; New York : North-Holland ; New York, N.Y., 144 1987, pp. 313-337. Gefen, D., and Straub, D.W. "Gender Differences in the Perception and Use of E-Mail: An Extension to the Technology Acceptance," MIS Quarterly (21:4), 1997, p 389. Grenci, R., and Todd, P. "Solutions-Driven Marketing," Communications of the ACM (45:3), 2002, pp 65-71. Griffin, D., and Tversky, A . "The Weighing of Evidence and the Determinants of Confidence," in: Heuristics and Biases, T. Gilovich, D. Griffin and D. Kahneman (eds.), Cambridge University Press, Cambridge, 2002, pp. 230-249. Haubl, G , and Murray, K . B . "Preference Construction and Persistence in Digital Marketplaces: The Role of Electronic Recommendation Agents," Journal of Consumer Psychology (13:1&2), 2003, p 75. Haubl, G , and Trifts, V. "Consumer Decision Making in Online Shopping Environments: the Effects of Interactive Decision Aids," Marketing Science (19:1), 2000, pp 189-200. Hogarth, R . M . Judgment and Choice: The Psychology of Decision, (2 ed.) Wiley, New York, 1987. Hostler, R.E., Yoon, V.Y., and Guimaraes, T. "Assessing the Impact of Internet Agent on End Users' Performance," Decision Support Systems (41:1), 2005. Hoyle, R.H., Kenny, D.A., and Hoyle, R.H. "Sample Size, Reliability, and Tests of Statistical Mediation," in: Statistical Strategies for Small Sample Research, R.H. Hoyle (ed.), Sage Publications, Inc., Thousand Oaks, California, 1999, pp. 195-222. Hsee, C.K., Loewenstein, G.F., Blount, S., and Bazerman, N . H . "Preference Reversals between Joint and Separate Evaluations of Options: a Review and Theoretical Analysis," Psychological bulletin (125), 1999, pp 576-590. Huber, O., and Seiser, G. "Accounting and Convincing: The Effect of Two Types of Justification on the Decision Process," Journal of Behavioral Decision Making (14:1), 2001, pp 69-85. Jarvenpaa, S.L. "The Effect of Task Demands and Graphical Format on Information Processing Strategies," Management Science (35:3), 1989, pp 285-303. Jiang, Z., and Benbasat, I. "An Experimental Investigation of Online Product Presentation Methods," Management Information Systems Quarterly), forthcoming. Joshi, A . "On Mobility and Agents," in: Mobile Networks and Computing, S. Rajasekaran, P. Pardalos and D.F. Hsu (eds.), American Mathematical Society, Providence, R. I., 2000, pp. 161-170. Kieras, D. "Using the Keystroke-Level Model to Estimate Execution Times," 2001, http://www.pitt.edu/~cmlewis/KSM.pdf, Last accessed on September 26, 2007. Kieras, D., and Helander, M.GL.T.K.a.P.P.V. "A Guide to GOMS Model Usability Evaluation Using NGOMSL," in: Handbook of Human-Computer Interaction, M . Helander, T.K. Landauer and P.V. Prabhu (eds.), Elsevier Science, 1997. Kim, D., and Benbasat, I. "The Effects of Trust-Assuring Arguments on Consumer Trust in Internet Stores: Application of Toulmin's Model of Argumentation," Information 145 Systems Research (17:3), 2006, pp 286-300. Kleinmuntz, D.N., and Schkade, D.A. "Information Displays and Decision Processes," Psychological Science (4), 1993, pp 221-227. Komiak, S.Y.X., and Benbasat, I. "The Effects of Personalization and Familiarity on Trust in and Adoption of Recommendation Agents," Management Information Systems Quarterly (30:4), 2006, pp 941-960. Komiak, S.Y.X., Wang, W , and Benbasat, I. "Trust Building in Virtual Salespersons versus in Human Salespersons: Similarities and Differences," E-Service Journal (3:3), 2004/2005. Kotteman, J.E., and Davis, F.D. "Decisional Conflict and User Acceptance of Multicriteria Decision-Making Aids," Decision Sciences (22:4), 1991, pp 918-926. Kotteman, J.E., Davis, F.D., and Remus, W.R. "Computer-Assisted Decision Making: Performance, Beliefs, and the Illusion of Control," Organizational Behavior and Human Decision Processes (57:1), 1994, pp 26-37. Koufaris, M . "Applying the Technology Acceptance Model and Flow Theory to Online Consumer Behavior," Information Systems Research (13:2), 2002, pp 205-223. Kunde, W. "Temporal Response-Effect Compatibility," Psychological Research (67:3), 2003, pp 153-159. Laudon, K.C. , and Laudon, J.P. Management Information Systems: Managing the Digital Firm, (9 ed.) Prentice Hall, Englewood Cliffs, New Jersey, 2006, p. 736. Lazarus, R.S. Emotion and Adaptation Oxford University Press, New York, 1991. Leavitt, N . "Recommendation Technology: Will It Boost E-Commerce?," Computer (39:5), 2006, pp 13-16. Lee, W., Liu, C , and Yeo, C. "Intelligent Agent-Based Systems for Personalized Recommendations in E-Commerce," Expert Systems with Applications (22:4), 2002, pp 275-284. Lee, Y.E., and Benbasat, I. "Interface Design for Mobile Commerce," Communications of the ACM (46:12), 2003, pp 48-52. Lee, Y.E., and Benbasat, I. "A Framework for the Study of Customer Interface Design for Mobile Commerce.," International Journal of Electronic Commerce (8:3), 2004, pp 79-102. Levin, A . M . , Levin, LP., and Heath, C E . "Product Category Dependent Consumer Preferences for Online and Offline Shopping Features and Their Influence on Multi-Channel Retail Alliances," Journal of Electronic Commerce Research (4:3), 2003, pp 85-93. Luce, M.F. "Choosing to Avoid: Coping with Negatively Emotion-Laden Consumer Decisions," Journal of Consumer Research (24:4), 1998, pp 409-433. Luce, M.F., Bettman, J.R., and Payne, J.W. "Choice Processing in Emotionally Difficult Decisions," Journal of Experimental Psychology: Learning, Memory, and Cognition (23:2), 1997, pp 384-405. 146 Luce, M.R, Bettman, J.R., and Payne, J.W. "Emotional Tradeoff Difficulty and Choice," Journal of Marketing Research (36:May), 1999, pp 143-159. Luce, M.F., Bettman, J.R., and Payne, J.W. "Coping with Unfavorable Attribute Values in Choice," Organizational behavior and human decision processes (81:22), 2000, pp 284-297. Luce, M.F., Bettman, J.R., Payne, J.W, and Bettman, J.R. Emotional Decisions: Tradeoff Difficulty and Coping in Consumer Choice The university of Chicago Press, Chicago, 2001. Maes, P., Guttman, R.H., and Moukas, A.G. "Agents that Buy and Sell," Communications of the ACM (42:3), 1999, pp 81-91. Mallalieu, L. "Consumer Perception of Salesperson Influence Strategies: An Examination of the Influence of Consumer Goals," Journal of Consumer Behavior (5:3), 2006, pp 257-268. McCabe, D.B., and Nowlis, S.M. "Information Integration across Online and Offline Shopping Environments and Its Effects on Consumers' Purchase Decisions," Experiential e-Commerce Conference, East Lansing, MI, 2001. Miller, B.N. , Albert, I., Lam, S.K., Konstan, J.A., and Riedl, J. "MovieLens Unplugged: Experiences with an Occasionally Connected Recommender System," International Conference on Intelligent User Interfaces 2003, Miami, Florida, 2003, pp. 263-266. Morin, S., Dube, L. , and Chebat, J. "The Role of Pleasant Music in Service Space: A Test of the Dual Model of Environmental Perception," Journal of Retailing (83:1), 2007, pp 115-130. Morris, S.A., and Marshall, T.E. "Perceived Control in Information Systems," Journal of Organizational and End User Computing (16:2), 2004, p 38. Murdock, B.B. "The Contributions of Hermann Ebbinhaus," Journal of Experimental Psychology: Learning, Memory, and Cognition (11:3), 1985, pp 469-471. Nowlis, S.M., and Simonson, I. "Attribute-Task Compatibility as a Determinant of Consumer Preference Reversals," Journal of Marketing Research (34:2), 1997, pp 205-218. O'Hara, K. , and Perry, M . "Shopping Anytime Anywhere," Conference on Human Factors in Computing Systems, Seattle, WA, 2001. O'Keefe, R .M. , and McEachern, T. "Web-Based Customer Decision Support Systems," Communications of the ACM (41:3), 1998, p 71. Olson, E X . , and Widing, R.E. "Are Interactive Decision Aids Better than Passive Decision Agents? A Comparison with Implications for Information Providers on the Internet," Journal of Interactive Marketing (16:2), 2002. Olson, J.R., and Olson. M , G. "The Growth of Cognitive Modeling in Human-Computer Interaction since GOMS," Human-Computer Interaction (5:2&3), 1990, pp 221-265. Pascoe, J., Ryan, N . , and Morse, D. "Human-Computer-Giraffe Interaction: HCI in the Field," the First Workshop on HCI with Mobile Devices, Glasgow, U K , 1998, p. 8. 147 Pascoe, J., Ryan, N . , and Morse, D. "Using While Moving: HCI Issues in Fieldwork Environment," Transactions on Computer-Human Interaction (7:3), 2000, pp 417-437. Payne, J.W, Bettman, J.R., and Johnson, E.J. "Adaptive Strategy Selection in Decision Making," Journal of Experimental Psychology: Learning, Memory, and Cognition (13:3), 1988, pp 534-552. Payne, J.W, Bettman, J.R., and Johnson, E.J. The Adaptive Decision-Maker Cambridge University Press, New York, 1993. Pereira, R.E. "Influence of Query-Based Decision Aids on Consumer Decision Making in Electronic Commerce," Information Resources Management Journal (14:1), 2001, p 31. Rosenthal, R., and Rosnow, R.L. Contrast Analysis: Focused Comparisons in the Analysis of Variance, Cambridge University Press, New York, 1985. Russo, J.E. "The Value of Unit Price Information," Journal of Marketing Research (14:2), 1977, pp 193-201. Russo, J.E., Staelin, R., Nolan, C.A., Russell, G.J., and Metcalf, B .L . "Nutrition Information in the Supermarket," Journal of Consumer Research (13:1), 1986, pp 48-70. Schwarz, N . , Higgins, E.T., and Sorrentino, R . M . "Feelings as Information: Informational and Motivational Functions of Affective States," in: Handbook of Motivation and Cognition: Foundations of Social Behavior, Anonymous (ed.), Guilford Press, New York, 1986. Shimp, T.A., Stuart, E.W., and Engle, R.W. "A Program of Classical Conditioning Experiments Testing Variations in the Conditional Stimulus and Context," Journal of Consumer Research (18:1), 1991, pp 1-12. Siegel, S., and Castellan, N.J. Nonparametric Statistics for the Behavioral Sciences, (2 ed.) McGraw-Hill New York, 1988. Singleton, R.A.J. , and Straits, B.C. Approaches to Social Research, Oxford University Press, New York, NY, 1999. Slovic, P. "From Shakespear to Simon: Speculation - and Some Evidence - about Man's Ability to Process Information," Oregon Research Institute Bulletin (12:3), 1972. Slovic, P. "The Construction of Preference," American Psychologist (50:5), 1995, pp 364-371. Slovic, P., Griffin, D., Tversky, A. , and Hogarth, R. "Compatibility Effects in Judgment and Choice," in: Insights in Decision Making: a Tribute to Hillel J. Einhorn, R. Hogarth (ed.), The University of Chicago Press, Chicago, 1990, pp. 5-27. Sobel, M.E. , and Leinhart, S. "Asymptotic Confidence Intervals for Indirect Effects in Structural Equation Models," in: Sociological Methodology, Anonymous (ed.), Jessey-Bass, San Francisco, 1982, pp. 290-312. Song, J., Jones, D., and Gudigantala, N . "The Effects of Incorporating Compensatory 148 Choice Strategies in Web-Based Consumer Decision Support Systems," Decision Support Systems (43:2), 2007, pp 359-374. Suh, K.S., and Lee, Y.E. "The Effects of Virual Reality on Consumer Learning: An Empirical Investigation," Management Information Systems Quarterly (29:4), 2005, pp 673-697. Svenson, O. "Process Descriptions of Decision Making," Organizational Behavior and Human Decision Processes (23:1), 1979, pp 86-112. Swaminathan, V. "The Impact of Recommendation Agents on Consumer Evaluation and Choice: the Moderating Role of Category Risk, Product Complexity, and Consumer Knowledge," Journal of Consumer Psychology (13:1&2), 2002, pp 93-101. Swearingen, K. , and Sinha, R. "Interaction Design for Recommender Systems," in: DIS 2002, London, 2002. Taylor, S., and Todd, P. "Assessing IT usage: The role of prior experience," MIS Quarterly (19:4), 1995, p 561. Tetlock, P.E., Kristel, O.V., Elson, B., and Green, M.C. "The Psychology of the Unthinkable: Taboo Trade-Offs, Forbidden Base Rates, and Heretical Counterfactuals," Journal of personality and social psychology (78:5), 2000, pp 853-870. Todd, P., and Benbasat, I. "An experimental investigation the impact of computer-based decision aids on the decision making process," 2 (2), 1991, pp 87-115. Todd, P., and Benbasat, I. "The Use of Information in Decision Making: An Experimental Investigation of the Impact of Computer-Based Decision Aids," Management Information Systems Quarterly (16:3), 1992, pp 373-393. Todd, P., and Benbasat, I. "The Influence of DSS on Choice Strategies: An Experimental Analysis of the Role of Cognitive Effort," Organizational behavior and human decision processes (60:1), 1994, pp 36-74. Todd, P., and Benbasat, I. "The Effects of Decision Support and Task Contingencies on Model Formulation: a Cognitive Perspective," Decision Support Systems (17), 1996, pp 241-252. Todd, P., and Benbasat, I. "Evaluating the Impact of DSS, Cognitive Effort, and Incentives on Strategy Selection," Information Systems Research (10:4), 1999, pp 356-374. Tversky, A. "Intransitivity of Preferences," Psychological Review (76:1), 1969, pp 31-48. Van der Heijden, H. "Mobile Decision Support for In-Store Purchase Decisions," Decision Support Systems (42:2), 2005. Van der Heijden, H. , and Sorensen, J.B. "The Mobile Decision Maker: Mobile Decision Aids, Task Complexity, and Decision Effectiveness," unpublished working paper, Copenhagen Business School, Fredericksburg, Denmark), 2002. Van der Heijden, H. , and Sorensen, L.S. "Observations on the Use of Mobile Decision Aids for Consumer Decision Making," International Journal of Management and 149 Decision Making (6:1), 2005, pp 5-15. Van Egeren, L . "Stress and Coping and Behavioral Organization," Psychosomatic Medicine (62:3), 2000, pp 451-460. Vijayasarathy, L.R., and Jones, J .M. "Do Internet Shopping Aids Make a Difference? An Empirical Investigation," Electronic Markets (11:1), 2001, pp 75-83. Wang, W , and Benbasat, I. "Trust In and Adoption of Online Recommendation Agents," Journal oftheAIS (6:3), 2005, pp 72-100. Wang, W., and Benbasat, I. "Impact of Explanations on Trust in and Adoption of Online Recommendation Agents," Journal of Management Information Systems (23:4), 2007, pp 219-249. Wickens, C D . Engineering Psychology and Human Performance Merrill, Columbus, Ohio, 1984. Widing, R.E., and Talarzyk, W.W. "Electronic Information Systems for Consumers: an Evaluation of Computer-Assisted Formats in Multiple Decision Environments," Journal of Marketing Research (30:2), 1993, pp 125-141. Wright, A .A . , and Lynch, J. "Communication Effects of Advertising versus Direct Experience when Both Search and Experience Attributes Are Present," Journal of Consumer Research (21:4), 1995, pp 708-718. Xiao, B., and Benbasat, I. "E-Commerce Product Recommendation Agents: Use, Characteristics, and Impact," Management Information Systems Quarterly (31:1), 2007, pp 137-209. Yates, F. "Contingency Table Involving Small Numbers and the Chi-Square Test," Journal of the Royal Statistical Society (Supplement) (1:2), 1934, pp 217-235. Youll, J., Morris, J., and Maes, P. "Impulse: Location-based Agent Assistance," Web Intelligence 2001,2001. Zeelenberg, M . "Anticipated Regret, Expected Feedback, and Behavioral Decision Making," Journal of Behavioral Decision Making (12:2), 1999, pp 93-106. 150 APPENDICES A.1 N A T U R A L G O M S L A N G U A G E A N A L Y S E S A . l . l Rationale for Conducting N G O M S L Analyses Natural GOMS Language (NGOMSL) is a cognitive modeling method created by Kieras (1997). Cognitive modeling methods are used to identify analytic models of human performance with systems, to analyze and list the sequence of steps required for users to perform a particular task within the system (Card, Moran and Newell 1983). Cognitive modeling methods were developed to overcome the soaring costs, time-lag, and poor representations of problem issues (e.g., users do not know what they need.) frequently found in empirical user testing (Kieras and Helander 1997). Cognitive modeling methods can be applied insofar as the user interface is specified in enough detail to dictate the sequence of actions required to perform the tasks, without the need for an implemented system or a mocked-up design, thereby accelerating data collection (Kieras and Helander 1997). Cognitive modeling methods, unlike purely empirical assessments, can also be used to capture the essence of the design in an inspectable representation (Kieras and Helander 1997). Given these strengths, cognitive modeling methods are often used as surrogates for actual user-data collection. GOMS was first proposed by Card, Moran, and Newell (1983). The acronym GOMS stands for Goals, Operators, Methods, and Selection Rules. Kieras (1997) describes GOMS as follows: Briefly, GOMS consists of descriptions of the Methods needed to accomplish specified Goals. The Methods are a series of steps consisting of Operators that the user performs. A Method may call for sub-Goals to be accomplished, so the Methods have a hierarchical structure. If there is more than one Method to accomplish a Goal, then Selection Rules choose the appropriate Method depending on the context. Describing the Goals, Operators, Methods, and Selection Rules for a set of tasks in a formal way constitutes doing a GOMS analysis, or constructing a GOMS model. 151 N G O M S L has significantly expanded GOMS from key-stroke-levels to higher-level program-like representations that the user must interact with to perform tasks with the system (Kieras and Helander 1997; Olson and Olson. M 1990). In addition to analyzing physical procedures (e.g., keystrokes), as traditional GOMS does, N G O M S L tracks both the contents and the resident duration of each piece of information stored in the user's working memory (Olson and Olson. M 1990). At the same time, N G O M S L enables the prediction of execution/learning time prior to system developments and the development of successful design insights, such as identifying which stage of activity takes the longest time to execute or produces the most errors (Kieras and Helander 1997; Olson and Olson. M 1990). Accordingly, I chose N G O M S L to analyze consumers' efforts to use RA-WAD and R A - E B A . A.1.2 Predicting Execution Time with N G O M S L Analyses According to Kieras (1997), execution time can be predicted by adding the time required to execute each N G O M S L statement and any operators included in the statements. Therefore, the estimated execution time is given by: Execution time = N G O M S L statement time + Primitive external operator time + Analyst-defined mental operator time + Waiting Time Execution time for each N G O M S L statement is 0.1 second per step (Kieras and Helander 1997). Waiting time is 0, given that, within the current study, retrieval of information is almost instantaneous and not the focus of my analyses. According to (Kieras and Helander 1997) (1997), waiting time can be ignored i f it is negligible and of no interest to analysts. Times for standard operators and mental operators are adopted from Kieras's (2001) study (see Table A . l ) . - INSERT [Table A . l ] H E R E --152 A . 1.3 N G O M S L Analyses I now present the N G O M S L steps required to use two RAs for a purchase decision task involving r number of attributes. Table A.2 shows the assumptions specific to this task. -- INSERT [Table A.2] H E R E --R A - W A D : RA-WAD involves 19*r+5 statements, (2*r + 1) P, (3*r+l) B B , (5*r+l) M , 2*r K, and 2*r H operators (see Table A.3). RA-WAD contains four higher-level goals: (1) users choose to employ the RA; (2) aided by the RA, users select the attributes to which they want to assign importance weights; (3) users assign importance weights to the attributes; and (4) users decide whether to select more attributes in the main menu. Users select the R A only once, but choose attributes, assign importance weights, and decide whether to continue repeatedly for each attribute. Therefore, the higher-level goals involve l+3*r statements. ~ INSERT [Table A.3] H E R E -The first goal, choosing the RA, involves 5 statements that include 1 P ( p o i n t in step 3), 1 BB ( c l i c k in step 4), 2 M (make up y o u r mind in step 1; f i n d in step 2). The second goal, choosing an attribute, consists of 5 statements that have embedded 1 P ( p o i n t in step 3), 1 BB ( c l i c k in step 4), and 2 M (make up y o u r mind in step 1; f i n d in step 2). These statements are repeated each time decision-makers choose an attribute and hence the number of statements and operators is multiplied by the number of attributes. Therefore, 5*r statements, l*r P, l*r B B , and 2*r M operators are necessary. The last goal, assigning importance weights, requires 10 statements that have 1 P ( p o i n t in step 3), 2 BB ( c l i c k in step 4; c l i c k submit button in step 9), 3 M ( n o t i c e in step 1; make up y o u r mind in step 2; v e r i f y in step 7), 1 K ( type 153 in step 6), and 2 H (move hands in step 5; move hands in step 8). A l l statements and embedded operators, except steps 2 and 6, are multiplied by the number of attributes since decision-makers assign importance weights to every attribute. For the last attribute, however, participants assign the remaining importance weight, rather than deciding upon the importance weights. Hence, one M operator, make up your mind in step 2, is repeated r-l times. One BB operator, Type in step 6, is repeated twice per attribute provided that most users type two keys (two-digit importance weights) for each attribute. Accordingly, 11 *r -1 statements, l*r P, 2*r B B , (3*r-l) M , 2*r K , and 2*r H operators are used. R A - E B A : R A - E B A needs 15*r+6 statements, (2*r+l) P, (2*r+l) B B , and (6*r+2) M operators (see Table A.4). R A - E B A contains three higher-level goals: (1) users choose the RA; (2) aided by the RA, users select the attributes on which they want to indicate the cutoff levels; (3) users choose the cutoff levels of the attributes. Finally, users decide whether or not to continue. Users choose the R A only once but repeatedly perform the rest for each attribute; thus 1 + 3*r statements are necessary. - INSERT [Table A.4] H E R E --As for RA-WAD, the choice of R A - E B A contains 5 statements, 1 P, 1 B B , and 2 M operators; the choice of attributes has 5*r statements, l*r P, l*r B B , arid 2*r M operators. The choice of cutoff levels requires 7 statements that include 1 P ( p o i n t in step 5), 1 BB ( c l i c k in step 6), and 4 M ( n o t i c e in step 1; n o t i c e in step 2; make up your mind in step 3; f i n d in step 4). Because users choose a cutoff level for each attribute, a total of 7*r statements, 1 *r P, 1 *r B B , and 4*r M operators are required. A.1.4 Summary Table A. 5 shows a summary of the N G O M S L analyses and the estimated execution times for RA-WAD and R A - E B A when the number of attributes (r) is 9 which is the number of attributes in the current study. RA-WAD and R A - E B A require similar numbers 154 of operators and similar execution times. - INSERT [Table A.5] HERE A.2 H I G H E M O T I O N D E C I S I O N T A S K Your ( ) has been driving a decent, upper-class vehicle appropriate for her age. Although it is not a luxurious car, this model is the regular pick of many consumer magazines in the reliable and safe vehicle category. In particular, it is known to have low breakdown rates of major parts, comprehensive safety features, and a high occupant survival rate reported in a nationwide crash test. Recently, she experienced serious financial problems, and can no longer afford to maintain this car. She needs to downgrade to another car. She has been searching for useful tips and information about how to find a used car. Since she doesn't know much about automobile parts and maintenance, she asked for a helping hand. She has an extra budget of $10,000 separate from her basic living expenses (e.g., food, housing, etc.) for this year. $10,000 will be allocated to: (1) purchasing a car and owning/maintaining it (i.e., Autoplan premiums, parking, gas, maintenance), and (2) emergency funds. In other words, this budget includes NOT O N L Y the car expenses BUT A L S O savings for unforeseen incidents: she wants to reserve some of the budget for any potential emergencies. In short, she strives to save money from this purchase. With her budget, she CANNOT purchase as good a vehicle as her current one. This is because most used cars in this price range ($10,000 or lower) are in average to poor condition. In other words, these vehicles are equipped with desirable and less desirable features. For instance, a car may have a reliable engine but a less reliable transmission. Recall that her current vehicle has low breakdown rates of many parts, comprehensive safety features, and high occupant survival rate. Therefore, she can only afford a car that is considerably inferior in many features. 156 A.3 L O W E M O T I O N DECIS ION T A S K Your ( ) has been driving a vehicle she bought from her friend who upgraded to a newer car a few years ago. This car is about 20 years old and has high breakdown rates of many parts, few safety features, and a. low occupant survival rate reported in a nationwide crash test. In the last few years, her financial status has improved substantially. She now wants to upgrade her vehicle. She has been searching for useful tips and information about how to find a decent used car. Since she doesn't know much about automobile parts and maintenance, she asked for a helping hand. She has an extra budget of $10,000 separate from her basic living expenses (e.g., food, housing, etc.) for this year. $10,000 will be allocated to: (1) purchasing a car and owning/maintaining it (i.e., Autoplan premiums, parking, gas, maintenance), and (2) emergency funds. In other words, this budget includes NOT O N L Y the car expenses BUT A L S O savings for unforeseen incidents, as she wants to reserve some of the budget for any potential emergencies. In short, she strives to save money from this purchase. Within her budget, she can definitely find a vehicle that is better than her current one. Most used cars in this price range ($10,000 or lower) are in average to poor condition. In other words, these vehicles are equipped with desirable and less desirable features. For instance, a car may have a reliable engine but a less reliable transmission. Recall that her current vehicle has high breakdown rates in many parts, few safety features, and a low occupant survival rate. Therefore, the she can afford a car that is considerably superior to her current one in many features. 157 A . 4 E X P E R I M E N T A L T A S K F O R T H E R A G R O U P S Shopping Task for a Family Member Your goal is to find the most suitable inkjet printer (NOT an all-in-one machine) as a birthday present for one of your family members at the shop created inside this room. Your family member does printing at home occasionally. S/he has little knowledge and experience of colour inkjet printers. Upon completion of the task, you will be asked to indicate your choice and to justify that choice. In your justification, you will describe: • What your family member requires of a colour inkjet printer; • What price range is acceptable for the gift; • Why you believe your chosen printer best satisfies his/her needs and stays within your price range for the gift. Shopping Task for a Friend Your goal is to find the most suitable inkjet printer (NOT an all-in-one machine) as a wedding gift for your best friend at the shop created inside this room. His/her work (or study) involves a lot of colour printing; hence s/he has substantial knowledge and experience of colour inkjet printers. Upon completion of the task, you will be asked to indicate your choice and to justify that choice. In your justification, you will describe: • What the friend of yours requires of a colour inkjet printer; • What price range is acceptable for the gift; • Why you believe your chosen printer best satisfies his/her needs and stays within your price range for the gift. 158 Two-Phased Shopping You will shop for a printer in two phases. The two phases reflect a typical consumer shopping pattern with the shopping advisor operating on a handheld device in a retail store. You make take as much time as you want to complete each phase. (1) Preliminary Product Viewing WITHOUT the Shopping Advisor: • Goal: Familiarize yourself with the printers sold in the store. • Walk around the store to view available printers. • When you are ready, exit the store and inform the assistant. (2) Making a Final Choice WITH the Help of the Shopping Advisor: • Goal: Choose one printer IN THE STORE using the shopping advisor. • Return to the store to make your final selection. • You may refer to the advisor at any time while making your final choice. • Ensure that you use the advisor at least once. • Write down the name of the printer you selected on a given sheet of paper. • Hand the sheet to the assistant when you exit the store. 159 A .5 E X P E R I M E N T A L T A S K S F O R T H E C O N T R O L G R O U P Shopping Task for a Family Member Your goal is to find the most suitable inkjet printer ( N O T an all-in-one machine) as a birthday present for one of your family members at the shop created inside this room. Your family member does printing at home occasionally. S/he has little knowledge and experience of colour inkjet printers. Upon completion of the task, you will be asked to indicate your choice and to justify that choice. In your justification, you will describe: • What the family member of yours requires of a colour inkjet printer; • What price range is acceptable for the gift; • Why you believe your chosen printer best satisfies his/her needs and stays within your price range for the gift. Shopping Task for a Friend Your goal is to find the most suitable inkjet printer ( N O T an all-in-one machine) as a wedding gift for your best friend at the shop created inside this room. His/her work (or study) involves a lot of colour printing; hence s/he has considerable knowledge and experience of colour inkjet printers. Upon completion of the task, you will be asked to indicate your choice and to justify that choice. In your justification, you will describe: • What the friend of yours requires of a colour inkjet printer; • What price range is acceptable for the gift; • Why you believe your chosen printer best satisfies his/her needs and stays within your price range for the gift. 160 Two-Phased Shopping You will shop for a printer in two phases. The two phases reflect a typical consumer shopping pattern in a retail store. You make take as much time as you want to complete each phase. (1) Preliminary Product Viewing: • Goal: Familiarize yourself with the printers sold in the store. • Walk around the store to view available printers. • When you are ready, exit the store and inform the assistant. (2) Making a Final Choice: • Goal: Choose one printer in the store. • Return to the store to make your final selection. • Write down the name of the printer you selected on a given sheet of paper. • Hand the sheet to the assistant when you exit the store. 161 A.6 T A B L E S Table A.1 Operators Adopted in the Current Study Operators User Activity Time (sec.) Standard operators (Attained from (Kieras 2001)) K Keystroke (Average non-secretarial typist) .28 T(n) Type a sequence of n characters on a keyboard n * K P Point with mouse to a target on the display 1.1 B Press or release mouse button .1 BB Push and release mouse button rapidly .2 H Home hands to keyboard or mouse .4 M Mental act of thinking or perception 1.2 Mental operators (Attained from (Kieras 2001)) Notice [value] Read specific instructions or options that require particular attention 1.2 Find [value] Locate an object in the display 1.2 Verify [value] Verify that some values are correctly input. 1.2 Make up your mind on [value] Decide on one out of several alternatives. 1.2 Table A.2 Assumptions of NGOMSL Analyses Overall Assumptions - Participants use RAs. - Users of RA-EBA and RA-WAD go through all attributes: they neither skip any attributes nor choose the same attribute more than once. - When choosing one out of a list of several options, users go through all options and then make a choice. - Hand starts and ends on mouse. - Average non-secretarial typists are assumed. Assumptions related to RA-WAD - Most users assign two-digit importance weights to attributes. - Users do not need to find the location of the input field, since the input field immediately follows the question, "How much importance would you like to assign to [attribute]? ". 162 Table A.3 NGOMSL Analyses of RA-WAD #of Statements Total # of Statements Type of Operators #of Operators Total # of Operators Method for goal Search f o r A l t e r n a t i v e s ' . ' Step 1 Accomplish g o a l : Choose recommendation agent. 1 1 Step 2 Accomplish g o a l : Choose a t t r i b u t e . 1 1*r Step 3 Accomplish g o a l : Assign importance weight. 1 1*r Step 6 Decide 1 1: i f no more a t t r i b u t e s l e f t , then r e t u r n w i t h goal accomplished. E l s e , go to 2 . 1 1*r Method for Goal Choose recommendation agent. Step 1 Make up your mind about the RA op t i o n . 1 1 M 1 1 Step 2 F i n d the RA opt i o n . 1 1 M 1 1 Step 3 Point to the RA option. 1 1 P 1 1 Step 4 C l i c k the RA opt i o n . 1 1 BB 1 1 Step 5 Return w i t h goal accomplished. 1 1 Method for goal Choose a t t r i b u t e . \'-r • • ' i- ... Step 1 Make up your mind about an a t t r i b u t e . 1 1*r M. 1 1*r Step 2 Find the a t t r i b u t e . 1 1*r M 1 1*r Step 3 Point to the a t t r i b u t e . 1 1*r P 1 1*r Step 4 C l i c k the a t t r i b u t e . 1 1*r BB 1 1*r Step 5 Return w i t h goal accomplished. 1 1*r Method for goal Assign importance weight. Step 1 Notice how many importance weights are l e f t . 1 1*r M 1 1*r Step 2 Make up your mind about how many weights to assign . 1 1*(r-1) M 1 1*(r-1) Step 3 Point to the f i e l d . 1 1*r P 1 1*r Step 4 C l i c k the f i e l d . 1 1*r BB 1 1*r Step 5 Move hands to keyboard. 1 1*r H 1 1*r Step 6 Type importance weight. 1 2*r K 1*2*r Step 7 V e r i f y that importance weights are typed c o r r e c t l y . 1 1*r M 1 1*r Step 8 Move hands back to mouse. 1 1*r H 1 1*r Step 9 C l i c k submit button. 1 1*r BB 1 1*r Step 10 Return w i t h goal accomplished. 1 1*r 11) Operators are underlined in statements. 163 Table A.4 NGOMSL Analyses of RA-EBA #of Statements Total # of Statements Type of Operators • #of Operators Total # of Operators Method for goal Choose the a l t e r n a t i v e . . Step 1 Accomplish g o a l : Choose recommendation agent. 1 1 Step 2 Accomplish g o a l : Choose a t t r i b u t e . 1 1*r Step 3 Accomplish g o a l : Choose c u t o f f l e v e l . 1 1*r Step 4 Decide: i f no more a t t r i b u t e s l e f t , then r e t u r n w i t h goal accomplished. E l s e , go to 2. 1 1*r Method for Goal Choose recommendation agent. Step 1 Make up your mind on the RA opt i o n . 1 1 M 1 1 Step 2 Find the RA op t i o n . 1 1 M 1 1 Step 3 Point to the RA op t i o n . 1 1 P 1 1 Step 4 C l i c k the RA op t i o n . 1 1 B B 1 1 Step 5 Return w i t h goal accomplished. 1 1 Method for goal Choose a t t r i b u t e . Step 1 Make up your mind on an a t t r i b u t e . 1 1*r M 1 1*r Step 2 Find the a t t r i b u t e . 1 1*r M 1 1*r Step 3 Point to the a t t r i b u t e . 1 1*r P 1 1*r Step 4 C l i c k the a t t r i b u t e . 1 1*r B B 1 1*r Step 5 Return w i t h goal accomplished. 1 1*r Method for goal Choose c u t o f f l e v e l . Step 1 Notice how many a l t e r n a t i v e s are l e f t . 1 1*r M 1 1*r Step 2 Notice how many a l t e r n a t i v e s are dropped by choosing each l e v e l . 1 1*r M 1 1*r Step 3 Make up your mind about which l e v e l to choose. 1 1*r M 1 1*r Step 4 Find the c u t o f f l e v e l . 1 1*r M 1 1*r Step 5 Point to the c u t o f f l e v e l . 1 1*r P 1 1*r Step 6 C l i c k the c u t o f f l e v e l . 1 1*r B B 1 1*r Step 7 Return with goal accomplished. 1 1*r 164 Table A.5 Summary of NGOMSL Analyses and the Estimated Execution Time RA-WAD R A - E B A N G O M S L Analyses (19T+5) (15*r+6) statements statements (2*r+1)P (2*r+ 1)P (2*r+1)BB (3*r+1)BB (6*r+2) M (5*r+1)M 2*rK 2*rH Estimated Execution Time (when r = 9) 111.54 seconds 106 seconds 1 2 r = number of attributes; each statement takes 0.1 sec to execute. 1 6 5 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0100651/manifest

Comment

Related Items