556 ieee transactions on systems, man, and cyberneticsâ€”part

556 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 35, NO. 4, JULY 2005

Interaction Effects of Consumers’ Product ClassKnowledge and Agent Search Strategy on Consumer

Decision Making in Electronic CommerceGene Rathnam

Abstract—This paper investigates the interaction effects be-tween the search strategy of software agents and consumers’product class knowledge in the context of consumers seeking topurchase cars on the Internet. The research design used was a2 4, between-groups, completely randomized, two-factor, facto-rial design. The independent variables that were manipulated wereproduct class knowledge (HIGH KNOWLEDGE, LOW KNOWL-EDGE) and agent search strategy [elimination by aspects (EBASTRATEGY), weighted average method (WAD STRATEGY),profile building (PROFILE STRATEGY), simple hypertext (HY-PERTEXT STRATEGY)]. The dependent variables that weremeasured were satisfaction with the decision process (SATIS-FACTION), confidence in the decision (CONFIDENCE), trust inthe agent’s recommendations (TRUST), propensity to purchase(PURCHASE), perceived cost savings (SAVINGS), and cognitivedecision effort (EFFORT). Significant differences were found inthe affective reactions of the subjects toward the agent/applicationdepending on the level of product class knowledge possessed bythe subjects. Subjects with high product class knowledge hadmore positive affective reactions toward agents/applications thatused the WAD and EBA strategies as compared to the PROFILEstrategy. Subjects with low product class knowledge had morepositive affective reactions to agents/applications that used thePROFILE strategy as compared to the EBA and WAD strategies.

Index Terms—Decision strategies, electronic commerce, productclass knowledge, search engines, software agents.

I. INTRODUCTION

MARKETING managers in the consumer goods industryface a new frontier of electronic information and com-

merce. Understanding buyer behavior in this new marketingchannel is crucial. Marketing managers would like to presentconsumers with information on which to base their decisions.The information presented has to be such that it allows con-sumers to make decisions and select products that best matchtheir tastes and needs [1]. Otherwise, consumers’ incentive toseek out information will be minimal [2]. Presenting such infor-mation is not simple. On the one hand, a vast amount of infor-mation could be relevant, even very relevant to some consumers.On the other hand, presenting superfluous information mightimpede consumers’ ability to make good decisions [1]. If con-sumers were predictable and all alike, presenting informationwould be simple—marketing managers could provide only the

Manuscript received October 1, 1999; revised July 30, 2004. This paper wasrecommended by Associate Editor Potter.

The author is with Satyam Consulting Services, Chicago, IL 60181 (e-mail:[email protected]).

Digital Object Identifier 10.1109/TSMCA.2005.850606

information that is deemed most relevant by all the consumers.However, because of the heterogeneity between consumers, andwithin consumers at different points in time, almost none of thepotentially available information is universally perceived as rel-evant. Rapid advancements in Internet technology have offereda solution to this dilemma in the form of computerized decisionaids that use software smart agents to provide an intelligent in-terface to the consumer.

The phenomenon of consumers purchasing products on theInternet is relatively new. In this business model, consumers se-lect items to purchase from electronic shopping malls by makingqueries to databases using software tools such as software smartagents. This has raised a host of interesting research issues thatneed to be investigated. The influence of electronic decision aidson satisfaction with the decision process, confidence in the de-cision, and the propensity to purchase has not been examinedpreviously and is of crucial importance.

The objective of this research is to understand why consumersreact the way they do to different forms of agent technologyand to identify mechanisms that will enable us to optimize thecognitive fit between an individual’s knowledge and expertiseand the decision strategy used by the software agent. To achievethis objective, four different decision environments were cre-ated, which varied the process used for filtering the information.The strategies used by the software agents for information fil-tering are discussed below.

1) The Weighted Average (WAD) strategy uses all the in-formation about the subject’s preferences and computesa weighted preference matching score for each alternativebased on the degree to which the attribute values of that al-ternative match the attribute values entered by the subjectas their preferred values and the preference weights allo-cated to each of the attributes by the subject. The alterna-tives are then sorted by the computed preference matchingscores and this sorted list is presented to the subject in de-scending order of the preference matching score. The sub-ject can browse this list and narrow their choice set by firstselecting the subset of alternatives he would like to con-sider further in their decision process and then systemati-cally eliminating alternatives from the choice set until hearrives at their final choice. The design of the applicationusing the WAD strategy is similar to that of the PersonalLogic web site (www.personalogic.com).

2) The Elimination By Aspects (EBA) strategy [3] obtainscutoff values from the decision maker on a set of productattributes. Alternatives that do not meet these specified

1083-4427/$20.00 © 2005 IEEE

Authorized licensed use limited to: Reza Mohammad Hashemi. Downloaded on September 25, 2009 at 16:47 from IEEE Xplore. Restrictions apply.

RATHNAM: INTERACTION EFFECTS OF CONSUMERS’ PRODUCT CLASS KNOWLEDGE AND AGENT SEARCH STRATEGY 557

criteria are eliminated from the choice set and the re-maining alternatives make up the reduced choice set.

3) The Profile Building (PROFILE) strategy matches theuser with “similar others” based on their demographicprofile (age, household income, gender, educationallevel). The agent deduces the user’s preferences bymatching him with a group of similar individuals basedon their demographic profile and expressed preferencesand predicts that their preferences in the product categoryunder consideration will also be similar to the user’s. Theagent will query the user for ratings on cars previouslyowned. Based on the demographic profile of the user, andthe ratings of previously owned cars, the agent will groupthe user with other individuals who responded similarlyand recommend the current car models that those indi-viduals have rated positively, inferring the user’s similarpreferences. Choices made by these similar others in theproduct category under consideration are used to form therecommended reduced choice set for the decision maker.The data to set up the different profiles was provided byIntelliChoice Inc., Campbell, CA. Six different profilecategories were created which included “Commuter,”“College Freshman,” “Executive,” “Soccer Mom,” “SportDriver,” and “Weekend Warrior.” The Amazon.com website employs such a profile-based strategy. The design ofthe application using the PROFILE strategy is similar tothat of the Firefly web site (www.firefly.net).

4) The Simple Hypertext (HYPERTEXT) strategy does notprovide any information filtering support to the user. Theuser is presented with a set of hypertext links to all the al-ternatives and is allowed to narrow the choice set at theirpace by sequentially eliminating alternatives from the ini-tial choice set.

II. THEORETICAL DEVELOPMENT OF HYPOTHESES

Many factors, including demographics and prior productclass experience, have been studied in an attempt to accountfor individual differences in consumers’ responses to a givenset of information [4]–[6]. Drawing on the information pro-cessing paradigm, this paper examines the effect of priorknowledge on information search behavior. A crucial elementin the information processing model of human behavior isinformation stored in memory—i.e., prior knowledge. Muchempirical evidence supports the view that prior knowledgeaffects information processing activities [7]–[11]. A number ofstudies have found a negative relationship between the amountof product experience and the amount of external search [6],[12]–[16]. An explanation for these results holds that experi-enced consumers perform more efficient information searchesbecause they know which attributes are most useful for dis-criminating between brands and can more quickly determinewhich alternatives are inferior. Knowledgeable consumerssubstitute internal search for external search, thus reducingtheir amount of external search for information. Furthermore,knowledgeable consumers search more efficiently. Knowledge-able consumers recognize product alternatives as belonging to

categories and subcategories [17]. Efficiency in informationsearch may occur in attribute selection as well as in alternativeselection. Specifically, highly knowledgeable consumers maysearch only those attributes that are useful for discriminatingamong alternatives. Knowledgeable consumers possess threetypes of knowledge that contribute to search efficiency by al-lowing a quicker elimination of unsuitable alternatives: criteriafor evaluating attributes, perceived covariance of attributes,and usage situation knowledge. Knowledge of criteria forevaluating attributes permits the individual to decide whetheran alternative is acceptable by allowing him to compare it toreference points stored in memory. Knowledge of attributecovariance allows inferences to be made about some attributeswithout external search. Usage situation knowledge leads toearlier—and possibly more accurate—categorization of alter-natives, based on their appropriateness for the intended usagesituation. And finally, in situations where known brands ratherthan hypothetical or unknown brands are used, knowledge ofthe attribute values of available brands is used as a substitutefor more effortful external search.

In providing agent services to consumers, intelligent agenttools depend on the accuracy of preference prediction to pro-vide benefits to the consumer [18]. The majority of the litera-ture examining intelligent agent-assisted electronic commerceassumes that consumers possess and are able to convey estab-lished preferences in the product category. An important prop-erty of the constructive viewpoint is that preferences will oftenbe extremely context-dependent. As such, an agent’s goal in achoice situation characterized by constructive preferences is notjust to inform, provide alternatives and uncover existing prefer-ences, but to help the consumer build preferences with the ulti-mate goal of aiding in choice.

An expert is someone who has acquired domain-specificknowledge through experience and training. This knowledgeresults in observable differences in cognitive processes butmay or may not lead to better performance in judgment anddecision making. This definition is consistent with most re-search reported in the behavioral decision-making literature. Toformulate a judgment, a decision maker must select, evaluate,and combine information that is available either internallyor externally. Empirical findings indicate that knowledgeabledecision makers are more selective in the information theyacquire [19], [20], are better able to acquire information ina less-structured environment [21], are more flexible in themanner in which they search for information [19], and agreemore than novices regarding what information is important[22]. A finding that has widespread support is that experts aremore confident in their decisions than are novices [23].

How problems are cognitively represented affects thestrategy by which problems are solved. Experts and noviceshave been found to exhibit differences in their problem-solvingstrategies. Experts categorize problems on the basis of solutionprocedures or underlying concepts, whereas novices tend tocategorize problems on the basis of surface features [11], [24].Because of this difference, experts use more efficient top-downor knowledge-based strategies, starting with known quantitiesto deduce unknowns [19], [24]. Should a solution path fail, thedecision maker can trace back a few steps and then proceed



again, a valuable process in determining what actions are ap-propriate. The bottom-up or means-end decision strategy usedby novices is not so practical. Decision makers start with thegoal and determine what conditions are necessary to achievethe goal. Because novices focus their attention on goals insteadof essential features of the task, learning is inhibited; they de-vote insufficient attention to acquiring valid schemas. The neteffect is that experts are credited with having better-developedprocedural knowledge [25]. If experts’ procedural knowledgeis appropriate for the task, they can encode and interpret in-formation more quickly than can novices [7], [19], [24]. Theyevoke a knowledge framework that is based on prior experiencethat expedites problem solving [26].

A view that addresses the effects of task characteristics onexpert judgments is that simple, well-structured domains donot give experts the opportunity to display their unique skills[19], [21]. Instead, general knowledge is sufficient to solve theproblem. There are benefits to expertise if the combination rulesare complex (environments that lend themselves to nonlinearcue use) and the task requires the decision maker to evaluateinputs [19].

Traditional laboratory tasks that organize the problem spaceby arranging decision alternatives and attributes in a matrix areinappropriate to use when studying expertise effects on judg-ment or choice processes [21]. In the real world, all the at-tributes and alternatives may not be readily available when thedecision maker first confronts the problem. Providing a matrixin the laboratory therefore could eliminate the special advan-tages that experts have in structuring problems. Domain exper-tise benefits consumers when the intended usage situation wascomplex, that is, when many attributes could be potentially rele-vant [21]. In contrast, there were no significant expertise effectswhen the intended usage situation was simple. Task characteris-tics, as opposed to individual differences, are the most importantdeterminants of behavior when the task is well-defined and un-ambiguous [27]. Individual differences take precedence as taskambiguity increases.

Problems can be characterized as structured, structurable, orunstructured for a particular problem solver at a given pointin time [28]. A problem is structured if the solver can readilyidentify a viable solution strategy. This might be the result ofprior experience solving the problem or of a problem beingstated in such a way that identifying a solution strategy istransparent. The application in which the search agent uses thePROFILE strategy would work well for structured problemssince the solution strategy is apparent to even the novice user.A problem is structurable if additional information wouldproduce a solution strategy or if the solver can reformulatethe problem into manageable subproblems, perhaps withthe aid of a structuring methodology. Although this class ofproblems is called structurable, only those with access to theright information or knowledge can structure it reliably. Theapplications in which the search agent uses the EBA strategy orthe WAD strategy would work well for structurable problemsbecause an expert with high product class knowledge would beable to add structure to the problem and formulate a solutionstrategy, whereas a novice with low product class knowledgewould not be able to add structure to the problem because of

TABLE ISUMMARY OF CONCEPTUAL FRAMEWORK

a lack of product class knowledge. For structurable problems,a complex but knowable solution strategy exists. A problem isinherently unstructured if a reliable solution strategy cannot belocated by any means. The HYPERTEXT strategy would betermed as unstructured because even experts cannot utilize theirproduct class knowledge to formulate a solution strategy to thisproblem. Problems that are unstructured are not necessarilyunsolvable. When faced with an unstructured problem, one isforced to employ less reliable methods and hope for the best[28].

When tasks are inherently unstructured (e.g., for the HY-PERTEXT strategy), even experts cannot apply known solutionstrategies. Instead, they must use heuristics, and their resultingjudgments are subject to all the biases associated with humanjudgment processes [29]. Expert performance under theseconditions is oftentimes poor [30]. Conversely, in domainsthat can be characterized as well structured (e.g., for the PRO-FILE strategy), general problem-solving knowledge could besufficient to solve the problem. Novices, therefore, could beexpected to induce reasonable problem-solving strategies onthe spot, so their performance might rival that of experts. It iswith tasks in the middle category, the set of ill-structured butstructurable problems (e.g., the EBA and WAD strategies), thatexperts would be expected to significantly outperform novices.Problems of this nature are information rich and require largeamounts of internal knowledge [31]. Using existing knowledge,experts can reformulate, decompose, and/or impose constraintsonto the problem, all of which reduce the size of the problemspace and hence its ambiguity [31]–[33]. The problem is nowinternally represented as one or more structured problems.Therefore, within the category of structurable problems, themore structure that is provided in the problem’s initial conditionthe better novices will do, decreasing the performance differ-ential between experts and novices. As problems become moreill-structured, the performance differential between experts andnovices also should decrease, because of experts’ decreasingability to structure the decision problem. The relationshipamong expertise, initial problem structure, and decision perfor-mance is illustrated in Table I.

To be useful, this conceptualization requires us to be ableto identify factors affecting a problem’s structure. Early re-searchers in this area propose that a problem’s initial structure isa function of how clearly specified are the goals, inputs, and/orallowable transformation rules-procedures used to integratedata inputs to reach a decision [33]–[36]. A list of problem



TABLE IIPROBLEM CHARACTERISTICS THAT AFFECT STRUCTURE

attributes that contribute to a problem’s level of structure hasbeen identified [28]. A set of problem characteristics is providedin Table II. For a problem to be (at least initially) ill-structured,at least one of these characteristics must be ambiguous.

To formulate a judgment about a prespecified criterion, a de-cision maker first must decide which inputs to use [37]. Em-pirical findings on information search point to dual benefits ofexpertise. On the one hand, expertise facilitates information pro-cessing, potentially increasing the amount of search; but on theother hand, expertise allows for more efficient searching, po-tentially decreasing search [21], [38]. Whether the facilitatingeffect or the efficiency effect dominates in a specific context de-pends partly on the demands of the task. If the decision makerhas access to large amounts of external information, much ofwhich is irrelevant or redundant, it is likely that experts will bemore selective in their use of available information than willnovices, because they can activate domain-specific schemas thatdirect attention to relevant information. In contrast, novices areless able to discern the diagnosticity of the available data and/orthe relationships among them.

Previous studies have found that experts find attributive state-ments such as those provided by the WAD and the EBA strate-gies informative, whereas novices consider benefit statementssuch as those provided by the PROFILE strategy informative[39], [40]. When attribute information is presented, as in theEBA and WAD strategies, experts have the knowledge to inferthe benefits implied by attribute information and are likely to bemotivated to make such inferences as a basis for judgments be-cause they perceive attribute information to be highly informa-tive. A technical attribute focus is likely to be effective becauseexperts are able to infer all of the related benefits and find thetechnical description to be more convincing [41]. By contrast, itseems unlikely that experts would process a benefits only mes-sage, such as that provided by the PROFILE strategy, in detail orbe persuaded by such an appeal. The absence of attribute infor-mation prevents experts from using their knowledge to evaluatethe claimed benefits. Without this information, experts may re-ject the communication as uninformative.

Novices are expected to respond differently than experts tothe attribute-oriented (EBA and WAD) and benefit-oriented(PROFILE) appeals. Physical features may be meaninglessto novices, so advertisements directed at them are structuredaround easily comprehended benefits [41]. This implies that

attribute-oriented information (EBA and WAD) would not beinformative and therefore not processed in detail by novices.By contrast, communications that focus on benefits (PROFILE)could be understood and perceived as informative by novices.

When subjects are using the EBA strategy, as the decisionprocess requires consideration of each product attribute and ex-pression of cutoff values for some of the attributes, it seemsreasonable to expect that it will be easier for individuals withhigh product class knowledge as compared to individuals withlow product class knowledge. Thus the EBA strategy will bepreferred by individuals with high product class knowledge ascompared to individuals with low product class knowledge. Ex-perts prefer to use noncompensatory strategies like eliminationby aspects and tend to process attribute information more ef-ficiently than novices [3], [42]. Experienced consumers knowwhich attributes are useful for distinguishing between optionsand may search only on those [21].

While knowledgeable consumers are well equipped toprocess attribute information, individuals with low productclass knowledge will likely find such a task more difficult.Attribute-oriented messages are found to be less informative tonovices [41], [43] as they do not process attribute informationas efficiently as experts. Given the added effort for individualswith low product class knowledge to process attribute infor-mation, it is expected that they will show negative affectivereactions to the agent/application when the agent is using theEBA strategy.

Hypothesis 1a: Subjects who have high product class knowl-edge will have more positive affective reactions to applicationsin which the search engines use the elimination by aspectfiltering strategy than subjects who have low product classknowledge.

As the Weighted Average strategy requires consideration ofeach product attribute and expression of preference with regardto each, it seems reasonable to expect that it will be easier forindividuals with high product class knowledge and thus will bepreferred by them.

Hypothesis 1b: Subjects who have high product class knowl-edge will have more positive affective reactions to applicationsin which the search engines use the weighted average filteringstrategy than subjects who have low product class knowledge.

Given their inability to process attribute information asefficiently as individuals with higher levels of product knowl-edge, consumers with low product class knowledge have beenfound to seek more summary information [21]. Because of thelower level of effort required, individuals with low productclass knowledge are predicted to prefer the PROFILE strategyas compared to the EBA or WAD strategies. Conversely, in-dividuals with high product class knowledge may base theirsearch only on product attributes [21]. The absence of attributeinformation in the PROFILE strategy prevents experts fromusing their knowledge to evaluate alternatives. Without thisopportunity experts may reject the communication as uninfor-mative [43] and will have negative affective reactions towardagents/applications that use the PROFILE strategy.

Thus, it is predicted that individuals with high product classknowledge will show less favorable response to agents usingthe PROFILE strategy than individuals with low product class



knowledge. The logic behind this is that experts are not per-mitted to use their attribute knowledge; rather, their preferencesare inferred by matching them with “similar others.” An under-lying factor in this negative response may also have to do withthe actual source of the recommendation—unidentified similarothers versus an internal agent algorithm driven by the user’sown preferences. Consumers trust their own opinions more thanthey trust the opinions of others [44], [45]. While other indi-viduals are thought to exhibit knowledge or reporting biases[46], making information provided by other sources ambiguous,self-generated information is less likely to be contaminated byknowledge or reporting biases.

Hypothesis 1c: Subjects who have low product class knowl-edge will have more positive affective reactions to applicationsin which the search engines use the profile building filteringstrategy than subjects who have high product class knowledge.

Since subjects with high product class knowledge have agreater ability to discriminate among alternatives and to elimi-nate undesirable alternatives from the choice set, it is expectedthat they will experience more positive affective reactions tothe HYPERTEXT strategy than subjects who have low productclass knowledge. However, because of the lack of agent supportin filtering the information, it is expected that subjects in boththe groups will experience significantly negative affective re-actions to the HYPERTEXT strategy as compared to the EBA,WAD, and PROFILE strategies.

III. DESCRIPTION OF VARIABLES

The scale items used to measure each of the constructs is pre-sented in Appendix A.

A. Independent Variables Being Manipulated

1) Agent Search Strategy (STRATEGY): Agent SearchStrategy refers to the search and decision strategy employed bythe agent in making recommendations to the user. The systemwas designed to have four treatment conditions, viz. WeightedAverage Strategy (WAD), Elimination By Aspects Strategy(EBA), Profile Building Strategy (PROFILE), and SimpleHypertext Strategy (HYPERTEXT).

2) Product Class Knowledge (KNOWLEDGE): Productclass knowledge refers to the knowledge about the product andthe familiarity with the product which the subject has.

B. Dependent Variables

1) Trust in the Agent’s Recommendations (TRUST): Trust inthe agent’s recommendations refers to the degree to which thesubject feels that the software agent has recommended alterna-tives to him that most closely match their preferences.

2) Propensity to Purchase (PURCHASE): Propensity toPurchase represents the subject’s perception that he wouldpurchase the selected alternative following the experiment if hewere going to make a purchase in that particular product class.

3) Satisfaction With the Decision Process (SATISFACTION):Satisfaction with the Decision Process represents the subject’ssubjective state of satisfaction with all aspects of the computer-ized decision process immediately after the decision has beenmade.

Fig. 1. The 2� 4, between-groups, two-factor experiment design.

4) Confidence in the Decision (CONFIDENCE): Confi-dence in the Decision refers to the confidence expressed by thesubject that he has selected the best alternative from the set offeasible alternatives.

5) Perceived Cost Savings (SAVINGS): Perceived Cost Sav-ings reflects the degree to which the subject feels that the use ofthe system has helped him to realize significant cost savings intheir purchase decision.

6) Cognitive Decision Effort (EFFORT): Cognitive Deci-sion Effort refers to the psychological costs of processing in-formation. This represents the ease with which the subject canperform the task of obtaining and processing the relevant infor-mation in order to enable him to arrive at their decision.

IV. RESEARCH METHODOLOGY FOR EXPERIMENT 1

A. Industry Selection

The product category “cars” was selected to maximize thelikelihood of a wide variation in product class knowledge forthe subject group used in the experiment. There are more than400 car models currently available in the U.S., and there existsa large number of attributes that customers typically use in theirselection of cars.

B. Experiment Design

The research design used was a 2 4, between-groups,completely randomized, two-factor, factorial design. The inde-pendent variables manipulated were product class knowledge(HIGH KNOWLEDGE, LOW KNOWLEDGE) and agentsearch strategy (EBA STRATEGY, WAD STRATEGY, PRO-FILE STRATEGY, HYPERTEXT STRATEGY). Subjects ineach group (HIGH KNOWLEDGE and LOW KNOWLEDGE)were randomly assigned to one of the four treatment conditions(EBA, WAD, PROFILE, HYPERTEXT). The experimentwas administered to 160 MBA students. Independent samplestesting was used. Twenty subjects were assigned to each cell.

The 2 4 design resulted in eight cells as is illustrated inFig. 1.

C. Factors Controlled in the Experiment

In order to test the hypothesized effects, the following factorswere controlled in the experiment.

1) Information Presentation Format and Graphics: This wasachieved through the use of a web site on a local server,designed specifically for this research.



2) Information Content of the Web Sites: For each alternativewhich the subject examined, the set of attributes for whichinformation was provided was the same. The attributesused included model year, body type, price, size (numberof passengers, number of doors, cargo capacity), technicalfeatures (engine type, transmission type, drive train type,brake type, number of cylinders, fuel efficiency, accelera-tion, braking distance, towing capacity), safety record andfeatures (antilock brakes, airbags, child safety locks, trac-tion control), maintenance costs, car manufacturer, andcountry of manufacture. In addition, a photograph of thecar model under examination was also placed on the website.

3) Download Time of the Web Pages: The download time ofthe web pages was constant for all the alternatives.

4) Number of Alternatives in the Feasible Set: The initialchoice set presented to the subjects was the same, regard-less of the treatment condition.

5) Choice Task: All the respondents were given the samechoice task, a setting in which they were asked to select acar to purchase from among the cars in the database.

6) Market Segment: The sample is based on a conveniencesample of MBA students. This represents a fairly homo-geneous market segment for the decision making task.

These controls limit the generalizability of the research, butare necessary to test the effects of interest.

D. Choice Environment

A laboratory experiment served as the vehicle for testing thehypotheses. Eight workstations in a behavioral decision labora-tory provided a controlled environment where subjects partici-pated in the experiment.

E. Choice Procedures

Subjects were first administered a preexperiment question-naire intended to determine initial knowledge in the product cat-egory. This included measures of familiarity with the productcategory and knowledge about the product category. The preex-periment questionnaire was also used to collect demographic in-formation from the subjects (e.g., their age, household income,education level, gender, etc.). The level of product class knowl-edge was manipulated by providing each of the subjects whowere identified following the initial questionnaire as belongingto the group “high product class knowledge” with training onproduct attributes and features. This consisted of a series ofscreens of terminology used to describe and assess attributesof the product. Subjects were able to click on any of the termsand receive additional information. Subjects in the “low productclass knowledge” group received no training on the product cat-egory. All subjects were then required to list from memory theattributes that they would use in their purchase decision for thisproduct. This provided a more objective measure of productclass knowledge. Consumers who have a consumption vocab-ulary are better able to develop and express preferences in aproduct category [47]. The subjects in each of the two groups“high” and “low” product class knowledge were then allocated

TABLE IIIRELIABILITY OF MEASURES USED

at random to one of the four treatment conditions. After com-pletion of the choice task, subjects were asked to rate the choicetask on the following dependent variables: cognitive effort, con-fidence in their decision, satisfaction with the decision process,propensity to purchase, perceived cost savings, and trust in theproduct alternatives recommended by the agent.

F. Manipulation Check

The initial subjective measure of product class knowledgeshowed some difference in product class knowledge betweensubjects in the “high product class knowledge” and “low productclass knowledge” groups. The initial measure of productclass knowledge was done by using a multiitem seven-pointLikert scale measure. Means for product class knowledgehigh knowledge low knowledge

were significantly different. Following the training task, thesubjects were asked to list the attributes they would considerin their purchase decision for cars. This was done to testsubjects’ objective knowledge about the product category. Thenumber of distinct, correct answers to the question was talliedto come up with an objective knowledge score. An independentscorer familiar with the product category and blind to thehypotheses being tested also graded the subjects’ responses.Interrater reliability was 0.93. The scores on this objectivemeasure of product class knowledge differed significantlybetween the “high product class knowledge” and the “lowproduct class knowledge” groups in the direction predicted,with subjects in the “high product class knowledge” groupbeing more knowledgeable about the terminology and choicefactors than those in the “low product class knowledge” group( mean high knowledge

mean low knowledge ; scores ranged from 0 to 11for the list of car attributes considered).

V. RESULTS FOR EXPERIMENT 1

Following extensive pretests, the measures used in the mainexperiment were found to have values for Cronbach rangingfrom 0.73 to 0.94. The results of the reliability tests are pre-sented in Table III. From this analysis, it can be concluded thatthe measures used had high reliability. Factor analysis of thedata indicated that the scale items loaded onto the constructsthey were a priori expected to load on. There were no cross-loadings. Furthermore, the number of factors that emerged wasidentical to those expected. Thus, there is statistical evidence to



TABLE IVCORRELATIONS BETWEEN THE CONSTRUCTS

support the claim that the scales have adequate unidimension-ality, convergent, and discriminant validity at the monomethodlevel of analysis. Hence it can be concluded that the measuresused had high validity. The results of the factor analysis are pre-sented in Table V(a) and (b). The measures used for the de-pendent variables SATISFACTION, CONFIDENCE, TRUST,PURCHASE, SAVINGS, and EFFORT were represented as themean-centered scores of the seven-point Likert scale items usedto measure these constructs. Correlation analysis shows thatproduct class knowledge is significantly correlated with sev-eral dependent variables, including satisfaction with the deci-sion process , trust in the agent’s recom-mendations and cognitive decision effort

. The results of the correlation analysisare presented in Table IV.

To rule out any potentially confounding effects, the samplemeans for computer familiarity were statistically comparedacross cells. Results of these tests indicated that each cellcontained subjects who, on average, had the same level ofcomputer familiarity. In addition, subjects were distributedapproximately equally across cells by gender. Hence, effectsidentified in the experimental data can be assigned with greatercertainty to the experimental variables under investigation.

Table VI presents the cell means obtained for each of the de-pendent variables. Table VII presents the summary of the sta-tistical analyses (F-values) for Experiment 1. The weighting forthe planned contrasts was consistent with the hypotheses beingtested [48]. The discussion of the results obtained for each ofthe dependent variables is presented below.

A. Satisfaction With the Decision Process (SATISFACTION)

A two-factor 2 4 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “satisfaction with thedecision process” (SATISFACTION). The factors used wereKNOWLEDGE and STRATEGY. This analysis yielded a sig-nificant overall interaction effect .The presence of interaction effects between KNOWLEDGEand STRATEGY for the dependent variable SATISFACTIONis clearly demonstrated in Fig. 2.

This was followed by an investigation of the simple effectsof the independent variable KNOWLEDGE on the depen-dent variable SATISFACTION. For the EBA STRATEGY,subjects with HIGH KNOWLEDGE experienced higher

TABLE VSUMMARY OF THE FACTOR ANALYSIS OF THE VARIABLES.

(a) PART 1. (b) PART 2

(a)

(b)

SATISFACTION than subjects with LOW KNOWLEDGE,. For the WAD STRATEGY,

subjects with HIGH KNOWLEDGE experienced higherSATISFACTION than subjects with LOW KNOWLEDGE,

. For the PROFILE STRATEGY,subjects with LOW KNOWLEDGE experienced higher SAT-ISFACTION than subjects with HIGH KNOWLEDGE,

. For the HYPERTEXTSTRATEGY, no significant difference in SATISFACTION wasdetected between subjects with HIGH KNOWLEDGE and sub-jects with LOW KNOWLEDGE, .Examination of the cell means shows clearly that for the



TABLE VICELL MEANS FOR EACH EMOTION AS A FUNCTION OF AGENT SEARCH

STRATEGY AND PRODUCT CLASS KNOWLEDGE FOR EXPERIMENT 1

TABLE VIISUMMARY OF THE STATISTICAL ANALYSES (F-VALUES) FOR EXPERIMENT 1

EBA STRATEGY and the WAD STRATEGY, SATISFAC-TION was higher for the HIGH KNOWLEDGE group ascompared to the LOW KNOWLEDGE group. For the PRO-FILE STRATEGY, SATISFACTION was higher for the LOWKNOWLEDGE group as compared to the HIGH KNOWL-EDGE group. An investigation of the simple effects of theindependent variable STRATEGY on the dependent variableSATISFACTION yielded significant results for both levels

Fig. 2. KNOWLEDGE� STRATEGY interaction for SATISFACTION.

of KNOWLEDGE. For the HIGH KNOWLEDGE group,. For the LOW KNOWLEDGE

group, .This was followed by a series of planned contrasts to inves-

tigate the effect of the independent variable STRATEGY onthe dependent variable SATISFACTION. The weighting forthe planned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY) and(HIGH KNOWLEDGE, WAD STRATEGY) with the com-bined data from the cells (HIGH KNOWLEDGE, PROFILESTRATEGY) and (HIGH KNOWLEDGE, HYPERTEXTSTRATEGY) indicated that subjects with HIGH KNOWL-EDGE experienced significantly higher SATISFACTION whenusing the EBA STRATEGY or the WAD STRATEGY ascompared to those who used the PROFILE STRATEGY or theHYPERTEXT STRATEGY, .A one way ANOVA conducted on the two cells (HIGHKNOWLEDGE, EBA STRATEGY) and (HIGH KNOWL-EDGE, WAD STRATEGY), combined for the purpose ofthe planned contrast, did not yield a significant difference,

. A one way ANOVA conducted onthe two cells (HIGH KNOWLEDGE, PROFILE STRATEGY)and (HIGH KNOWLEDGE, HYPERTEXT STRATEGY),combined for the purpose of the planned contrast, did notyield a significant difference, .A planned contrast comparing the data from the cell (LOWKNOWLEDGE, PROFILE STRATEGY) with the combineddata from the cells (LOW KNOWLEDGE, EBA STRATEGY),(LOW KNOWLEDGE, WAD STRATEGY), and (LOWKNOWLEDGE, HYPERTEXT STRATEGY) indicated thatsubjects with LOW KNOWLEDGE experienced significantlyhigher SATISFACTION when using PROFILE STRATEGYas compared to those who used the EBA STRATEGY,the WAD STRATEGY, or the HYPERTEXT STRATEGY,

. A one way ANOVA whichwas conducted on the three cells (LOW KNOWLEDGE, EBASTRATEGY), (LOW KNOWLEDGE, WAD STRATEGY), and(LOW KNOWLEDGE, HYPERTEXT STRATEGY) whichwere combined for the purpose of the planned contrast didnot yield a significant difference, .Examination of the cell means shows clearly that the HIGH



Fig. 3. KNOWLEDGE� STRATEGY interaction for CONFIDENCE.

KNOWLEDGE group experienced higher SATISFACTIONwhen using the EBA STRATEGY or the WAD STRATEGYand the LOW KNOWLEDGE group experienced higher SAT-ISFACTION when using the PROFILE STRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variableSATISFACTION. The weighting for the complex interactioncontrast was consistent with the hypothesis being tested [48].A complex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result

. From the above-mentionedseries of statistical analyses, it was concluded that there existsa strong interaction effect between the independent variablesKNOWLEDGE and STRATEGY on the dependent variableSATISFACTION. Subjects with HIGH KNOWLEDGE expe-rience significantly higher SATISFACTION when using theEBA STRATEGY or the WAD STRATEGY. Subjects withLOW KNOWLEDGE experience significantly higher SATIS-FACTION when using the PROFILE STRATEGY.

B. Confidence in the Decision (CONFIDENCE)

A two-factor 2 4 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “confidence in thedecision” (CONFIDENCE). The factors used were KNOWL-EDGE and STRATEGY. This analysis yielded a significantoverall interaction effect . Thepresence of interaction effects between KNOWLEDGE andSTRATEGY for the dependent variable CONFIDENCE isclearly demonstrated in Fig. 3.

This was followed by an investigation of the simple effectsof the independent variable KNOWLEDGE on the depen-dent variable CONFIDENCE. For the EBA STRATEGY,subjects with HIGH KNOWLEDGE experienced higherCONFIDENCE than subjects with LOW KNOWLEDGE,

. For the WAD STRATEGY,subjects with HIGH KNOWLEDGE experienced higherCONFIDENCE than subjects with LOW KNOWLEDGE,

. For the PROFILE STRATEGY,subjects with LOW KNOWLEDGE experienced higherCONFIDENCE than subjects with HIGH KNOWLEDGE,

. For the HYPERTEXTSTRATEGY, no significant difference in CONFIDENCE wasdetected between subjects with HIGH KNOWLEDGE and sub-jects with LOW KNOWLEDGE, .Examination of the cell means shows clearly that for theEBA STRATEGY and the WAD STRATEGY, CONFI-DENCE was higher for the HIGH KNOWLEDGE group ascompared to the LOW KNOWLEDGE group. For the PRO-FILE STRATEGY, CONFIDENCE was higher for the LOWKNOWLEDGE group as compared to the HIGH KNOWL-EDGE group. An investigation of the simple effects of theindependent variable STRATEGY on the dependent variableCONFIDENCE yielded significant results for both levelsof KNOWLEDGE. For the HIGH KNOWLEDGE group,

. For the LOW KNOWLEDGEgroup, .

This was followed by a series of planned contrasts to in-vestigate the effect of the independent variable STRATEGYon the dependent variable CONFIDENCE. The weighting forthe planned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY) and(HIGH KNOWLEDGE, WAD STRATEGY) with the com-bined data from the cells (HIGH KNOWLEDGE, PROFILESTRATEGY) and (HIGH KNOWLEDGE, HYPERTEXTSTRATEGY) indicated that subjects with HIGH KNOWL-EDGE experienced significantly higher CONFIDENCE whenusing the EBA STRATEGY or the WAD STRATEGY ascompared to those who used the PROFILE STRATEGY or theHYPERTEXT STRATEGY, . Aone way ANOVA which was conducted on the two cells (HIGHKNOWLEDGE, EBA STRATEGY) and (HIGH KNOWL-EDGE, WAD STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A one way ANOVAwhich was conducted on the two cells (HIGH KNOWL-EDGE, PROFILE STRATEGY) and (HIGH KNOWLEDGE,HYPERTEXT STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A planned contrastcomparing the data from the cell (LOW KNOWLEDGE,PROFILE STRATEGY) with the combined data from the cells(LOW KNOWLEDGE, EBA STRATEGY), (LOW KNOWL-EDGE, WAD STRATEGY), and (LOW KNOWLEDGE,HYPERTEXT STRATEGY) indicated that subjects with LOWKNOWLEDGE experienced significantly higher CONFI-DENCE when using PROFILE STRATEGY as compared tothose who used the EBA STRATEGY, the WAD STRATEGY, orthe HYPERTEXT STRATEGY, . Aone way ANOVA which was conducted on the three cells (LOWKNOWLEDGE, EBA STRATEGY), (LOW KNOWLEDGE,WAD STRATEGY), and (LOW KNOWLEDGE, HYPER-TEXT STRATEGY) which were combined for the purposeof the planned contrast did not yield a significant difference,

. Examination of the cell means



Fig. 4. KNOWLEDGE� STRATEGY interaction for TRUST.

shows clearly that the HIGH KNOWLEDGE group experiencedhigher CONFIDENCE when using the EBA STRATEGY orthe WAD STRATEGY and the LOW KNOWLEDGE groupexperienced higher CONFIDENCE when using the PROFILESTRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variableCONFIDENCE. The weighting for the complex interactioncontrast was consistent with the hypothesis being tested [48].A complex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result

. From the above-mentionedseries of statistical analyses, it was concluded that there existsa strong interaction effect between the independent variablesKNOWLEDGE and STRATEGY on the dependent variableCONFIDENCE. Subjects with HIGH KNOWLEDGE expe-rience significantly higher CONFIDENCE when using theEBA STRATEGY or the WAD STRATEGY. Subjects withLOW KNOWLEDGE experience significantly higher CONFI-DENCE when using the PROFILE STRATEGY.

C. Trust in the Agent’s Recommendations (TRUST)

A two-factor 2 3 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “trust in the agent’srecommendations” (TRUST). The factors used were KNOWL-EDGE and STRATEGY. This analysis yielded a significantoverall interaction effect . Thepresence of interaction effects between KNOWLEDGE andSTRATEGY for the dependent variable TRUST is clearlydemonstrated in Fig. 4.

This was followed by an investigation of the simple effects ofthe independent variable KNOWLEDGE on the dependent vari-able TRUST. For the EBA STRATEGY, subjects with HIGHKNOWLEDGE experienced higher TRUST than subjectswith LOW KNOWLEDGE, .For the WAD STRATEGY, subjects with HIGH KNOWL-EDGE experienced higher TRUST than subjects with LOW

KNOWLEDGE, . For thePROFILE STRATEGY, subjects with LOW KNOWLEDGEexperienced higher TRUST than subjects with HIGH KNOWL-EDGE, . Examination of thecell means shows clearly that for the EBA STRATEGY andthe WAD STRATEGY, TRUST was higher for the HIGHKNOWLEDGE group as compared to the LOW KNOWL-EDGE group. For the PROFILE STRATEGY, TRUST washigher for the LOW KNOWLEDGE group as compared tothe HIGH KNOWLEDGE group. An investigation of thesimple effects of the independent variable STRATEGY on thedependent variable TRUST yielded significant results for bothlevels of KNOWLEDGE. For the HIGH KNOWLEDGE group,


This was followed by a series of planned contrasts to in-vestigate the effect of the independent variable STRATEGYon the dependent variable TRUST. The weighting for theplanned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY)and (HIGH KNOWLEDGE, WAD STRATEGY) with the datafrom the cell (HIGH KNOWLEDGE, PROFILE STRATEGY)indicated that subjects with HIGH KNOWLEDGE experiencedsignificantly higher TRUST when using the EBA STRATEGYor the WAD STRATEGY as compared to those who usedthe PROFILE STRATEGY, .A one way ANOVA which was conducted on the two cells(HIGH KNOWLEDGE, EBA STRATEGY) and (HIGHKNOWLEDGE, WAD STRATEGY) which were combinedfor the purpose of the planned contrast did not yield a sig-nificant difference, . A plannedcontrast comparing the data from the cell (LOW KNOWL-EDGE, PROFILE STRATEGY) with the combined datafrom the cells (LOW KNOWLEDGE, EBA STRATEGY),and (LOW KNOWLEDGE, WAD STRATEGY) indicatedthat subjects with LOW KNOWLEDGE experienced signif-icantly higher TRUST when using PROFILE STRATEGYas compared to those who used the EBA STRATEGY or theWAD STRATEGY, . A oneway ANOVA which was conducted on the two cells (LOWKNOWLEDGE, EBA STRATEGY) and (LOW KNOWL-EDGE, WAD STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . Examination of thecell means shows clearly that the HIGH KNOWLEDGE groupexperienced higher TRUST when using the EBA STRATEGYor the WAD STRATEGY and the LOW KNOWLEDGEgroup experienced higher TRUST when using the PROFILESTRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variableTRUST. The weighting for the complex interaction con-trast was consistent with the hypothesis being tested [48]. Acomplex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result



Fig. 5. KNOWLEDGE� STRATEGY interaction for PURCHASE.

. From the above-mentionedseries of statistical analyses, it was concluded that there existsa strong interaction effect between the independent variablesKNOWLEDGE and STRATEGY on the dependent variableTRUST. Subjects with HIGH KNOWLEDGE experience sig-nificantly higher TRUST when using the EBA STRATEGY orthe WAD STRATEGY. Subjects with LOW KNOWLEDGEexperience significantly higher TRUST when using the PRO-FILE STRATEGY.

D. Propensity to Purchase (PURCHASE)

A two-factor 2 4 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “propensity to pur-chase” (PURCHASE). The factors used were KNOWLEDGEand STRATEGY. This analysis yielded a significant overallinteraction effect . The presenceof interaction effects between KNOWLEDGE and STRATEGYfor the dependent variable PURCHASE is clearly demonstratedin Fig. 5.

This was followed by an investigation of the simple ef-fects of the independent variable KNOWLEDGE on thedependent variable PURCHASE. For the EBA STRATEGY,subjects with HIGH KNOWLEDGE experienced higherPURCHASE than subjects with LOW KNOWLEDGE,

. For the WAD STRATEGY,subjects with HIGH KNOWLEDGE experienced higherPURCHASE than subjects with LOW KNOWLEDGE,

. For the PROFILE STRATEGY,subjects with LOW KNOWLEDGE experienced higherPURCHASE than subjects with HIGH KNOWLEDGE,

. For the HYPERTEXTSTRATEGY, no significant difference in PURCHASE wasdetected between subjects with HIGH KNOWLEDGE and sub-jects with LOW KNOWLEDGE, .Examination of the cell means shows clearly that for the EBASTRATEGY and the WAD STRATEGY, PURCHASE washigher for the HIGH KNOWLEDGE group as compared to theLOW KNOWLEDGE group. For the PROFILE STRATEGY,PURCHASE was higher for the LOW KNOWLEDGE group ascompared to the HIGH KNOWLEDGE group. An investigation

of the simple effects of the independent variable STRATEGYon the dependent variable PURCHASE yielded significantresults for both levels of KNOWLEDGE. For the HIGHKNOWLEDGE group, . For theLOW KNOWLEDGE group, .

This was followed by a series of planned contrasts to in-vestigate the effect of the independent variable STRATEGYon the dependent variable PURCHASE. The weighting for theplanned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY) and(HIGH KNOWLEDGE, WAD STRATEGY) with the com-bined data from the cells (HIGH KNOWLEDGE, PROFILESTRATEGY) and (HIGH KNOWLEDGE, HYPERTEXTSTRATEGY) indicated that subjects with HIGH KNOWL-EDGE experienced significantly higher PURCHASE whenusing the EBA STRATEGY or the WAD STRATEGY ascompared to those who used the PROFILE STRATEGY or theHYPERTEXT STRATEGY, . Aone way ANOVA which was conducted on the two cells (HIGHKNOWLEDGE, EBA STRATEGY) and (HIGH KNOWL-EDGE, WAD STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A one way ANOVAwhich was conducted on the two cells (HIGH KNOWL-EDGE, PROFILE STRATEGY) and (HIGH KNOWLEDGE,HYPERTEXT STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A planned contrastcomparing the data from the cell (LOW KNOWLEDGE,PROFILE STRATEGY) with the combined data from the cells(LOW KNOWLEDGE, EBA STRATEGY), (LOW KNOWL-EDGE, WAD STRATEGY), and (LOW KNOWLEDGE,HYPERTEXT STRATEGY) indicated that subjects with LOWKNOWLEDGE experienced significantly higher PURCHASEwhen using PROFILE STRATEGY as compared to those whoused the EBA STRATEGY, the WAD STRATEGY, or the HY-PERTEXT STRATEGY, . A oneway ANOVA which was conducted on the three cells (LOWKNOWLEDGE, EBA STRATEGY), (LOW KNOWLEDGE,WAD STRATEGY), and (LOW KNOWLEDGE, HYPER-TEXT STRATEGY) which were combined for the purposeof the planned contrast did not yield a significant difference,

. Examination of the cell meansshows clearly that the HIGH KNOWLEDGE group experi-enced higher PURCHASE when using the EBA STRATEGYor the WAD STRATEGY and the LOW KNOWLEDGE groupexperienced higher PURCHASE when using the PROFILESTRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variablePURCHASE. The weighting for the complex interaction con-trast was consistent with the hypothesis being tested [48]. Acomplex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result

. From the above-mentioned



Fig. 6. KNOWLEDGE� STRATEGY interaction for SAVINGS.

series of statistical analyses, it was concluded that there existsa strong interaction effect between the independent variablesKNOWLEDGE and STRATEGY on the dependent variablePURCHASE. Subjects with HIGH KNOWLEDGE experi-ence significantly higher PURCHASE when using the EBASTRATEGY or the WAD STRATEGY. Subjects with LOWKNOWLEDGE experience significantly higher PURCHASEwhen using the PROFILE STRATEGY.

E. Perceived Cost Savings (SAVINGS)

A two-factor 2 4 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “perceived costsavings” (SAVINGS). The factors used were KNOWLEDGEand STRATEGY. This analysis yielded a significant overallinteraction effect . The presenceof interaction effects between KNOWLEDGE and STRATEGYfor the dependent variable SAVINGS is clearly demonstratedin Fig. 6.

This was followed by an investigation of the simple effects ofthe independent variable KNOWLEDGE on the dependent vari-able SAVINGS. For the EBA STRATEGY, subjects with HIGHKNOWLEDGE experienced higher SAVINGS than subjectswith LOW KNOWLEDGE, .For the WAD STRATEGY, subjects with HIGH KNOWL-EDGE experienced higher SAVINGS than subjects withLOW KNOWLEDGE, . Forthe PROFILE STRATEGY, subjects with LOW KNOWL-EDGE experienced higher SAVINGS than subjects withHIGH KNOWLEDGE, . Forthe HYPERTEXT STRATEGY, no significant differencein SAVINGS was detected between subjects with HIGHKNOWLEDGE and subjects with LOW KNOWLEDGE,

. Examination of the cell meansshows clearly that for the EBA STRATEGY and the WADSTRATEGY, SAVINGS was higher for the HIGH KNOWL-EDGE group as compared to the LOW KNOWLEDGE group.For the PROFILE STRATEGY, SAVINGS was higher forthe LOW KNOWLEDGE group as compared to the HIGHKNOWLEDGE group. An investigation of the simple effects

of the independent variable STRATEGY on the dependentvariable SAVINGS yielded significant results for both levelsof KNOWLEDGE. For the HIGH KNOWLEDGE group,


This was followed by a series of planned contrasts to in-vestigate the effect of the independent variable STRATEGYon the dependent variable SAVINGS. The weighting for theplanned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY) and(HIGH KNOWLEDGE, WAD STRATEGY) with the com-bined data from the cells (HIGH KNOWLEDGE, PROFILESTRATEGY) and (HIGH KNOWLEDGE, HYPERTEXTSTRATEGY) indicated that subjects with HIGH KNOWL-EDGE experienced significantly higher SAVINGS when usingthe EBA STRATEGY or the WAD STRATEGY as comparedto those who used the PROFILE STRATEGY or the HYPER-TEXT STRATEGY, . A oneway ANOVA which was conducted on the two cells (HIGHKNOWLEDGE, EBA STRATEGY) and (HIGH KNOWL-EDGE, WAD STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A one way ANOVAwhich was conducted on the two cells (HIGH KNOWL-EDGE, PROFILE STRATEGY) and (HIGH KNOWLEDGE,HYPERTEXT STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A planned contrastcomparing the data from the cell (LOW KNOWLEDGE,PROFILE STRATEGY) with the combined data from the cells(LOW KNOWLEDGE, EBA STRATEGY), (LOW KNOWL-EDGE, WAD STRATEGY), and (LOW KNOWLEDGE,HYPERTEXT STRATEGY) indicated that subjects with LOWKNOWLEDGE experienced significantly higher SAVINGSwhen using PROFILE STRATEGY as compared to those whoused the EBA STRATEGY, the WAD STRATEGY, or the HY-PERTEXT STRATEGY, . A oneway ANOVA which was conducted on the three cells (LOWKNOWLEDGE, EBA STRATEGY), (LOW KNOWLEDGE,WAD STRATEGY), and (LOW KNOWLEDGE, HYPER-TEXT STRATEGY) which were combined for the purposeof the planned contrast did not yield a significant difference,

. Examination of the cell meansshows clearly that the HIGH KNOWLEDGE group experi-enced higher SAVINGS when using the EBA STRATEGY orthe WAD STRATEGY and the LOW KNOWLEDGE groupexperienced higher SAVINGS when using the PROFILESTRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variableSAVINGS. The weighting for the complex interaction con-trast was consistent with the hypothesis being tested [48]. Acomplex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result

. From the above-mentioned



Fig. 7. KNOWLEDGE� STRATEGY interaction for EFFORT.

series of statistical analyses, it was concluded that thereexists a strong interaction effect between the independentvariables KNOWLEDGE and STRATEGY on the dependentvariable SAVINGS. Subjects with HIGH KNOWLEDGE ex-perience significantly higher SAVINGS when using the EBASTRATEGY or the WAD STRATEGY. Subjects with LOWKNOWLEDGE experience significantly higher SAVINGSwhen using the PROFILE STRATEGY.

F. Cognitive Decision Effort (EFFORT)

A two-factor 2 4 ANOVA was performed to test the in-teraction effects of the independent variables “product classknowledge” (KNOWLEDGE) and “agent search strategy”(STRATEGY) on the dependent variable “cognitive decisioneffort” (EFFORT). The factors used were KNOWLEDGE andSTRATEGY. This analysis yielded a significant overall inter-action effect . The presence ofinteraction effects between KNOWLEDGE and STRATEGYfor the dependent variable EFFORT is clearly demonstrated inFig. 7.

This was followed by an investigation of the simple effects ofthe independent variable KNOWLEDGE on the dependent vari-able EFFORT. For the EBA STRATEGY, subjects with HIGHKNOWLEDGE experienced lower EFFORT than subjectswith LOW KNOWLEDGE, .For the WAD STRATEGY, subjects with HIGH KNOWL-EDGE experienced lower EFFORT than subjects with LOWKNOWLEDGE, . For thePROFILE STRATEGY, subjects with LOW KNOWLEDGEexperienced lower EFFORT than subjects with HIGH KNOWL-EDGE, . For the HYPERTEXTSTRATEGY, no significant difference in EFFORT was detectedbetween subjects with HIGH KNOWLEDGE and subjects withLOW KNOWLEDGE, . Ex-amination of the cell means shows clearly that for the EBASTRATEGY and the WAD STRATEGY, EFFORT was lowerfor the HIGH KNOWLEDGE group as compared to theLOW KNOWLEDGE group. For the PROFILE STRATEGY,EFFORT was lower for the LOW KNOWLEDGE group ascompared to the HIGH KNOWLEDGE group. An investigationof the simple effects of the independent variable STRATEGY

on the dependent variable EFFORT yielded significant resultsfor both levels of KNOWLEDGE. For the HIGH KNOWL-EDGE group, . For the LOWKNOWLEDGE group, .

This was followed by a series of planned contrasts to in-vestigate the effect of the independent variable STRATEGYon the dependent variable EFFORT. The weighting for theplanned contrasts was consistent with the hypotheses beingtested [48]. A planned contrast comparing the combined datafrom the cells (HIGH KNOWLEDGE, EBA STRATEGY) and(HIGH KNOWLEDGE, WAD STRATEGY) with the com-bined data from the cells (HIGH KNOWLEDGE, PROFILESTRATEGY) and (HIGH KNOWLEDGE, HYPERTEXTSTRATEGY) indicated that subjects with HIGH KNOWL-EDGE experienced significantly lower EFFORT when usingthe EBA STRATEGY or the WAD STRATEGY as comparedto those who used the PROFILE STRATEGY or the HYPER-TEXT STRATEGY, . A oneway ANOVA which was conducted on the two cells (HIGHKNOWLEDGE, EBA STRATEGY) and (HIGH KNOWL-EDGE, WAD STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A one way ANOVAwhich was conducted on the two cells (HIGH KNOWL-EDGE, PROFILE STRATEGY) and (HIGH KNOWLEDGE,HYPERTEXT STRATEGY) which were combined for thepurpose of the planned contrast did not yield a significantdifference, . A planned contrastcomparing the data from the cell (LOW KNOWLEDGE,PROFILE STRATEGY) with the combined data from the cells(LOW KNOWLEDGE, EBA STRATEGY), (LOW KNOWL-EDGE, WAD STRATEGY), and (LOW KNOWLEDGE,HYPERTEXT STRATEGY) indicated that subjects with LOWKNOWLEDGE experienced significantly lower EFFORT whenusing PROFILE STRATEGY as compared to those who usedthe EBA STRATEGY, the WAD STRATEGY, or the HYPER-TEXT STRATEGY, . A oneway ANOVA which was conducted on the three cells (LOWKNOWLEDGE, EBA STRATEGY), (LOW KNOWLEDGE,WAD STRATEGY), and (LOW KNOWLEDGE, HYPER-TEXT STRATEGY) which were combined for the purposeof the planned contrast did not yield a significant difference,

. Examination of the cell meansshows clearly that the HIGH KNOWLEDGE group experiencedlower EFFORT when using the EBA STRATEGY or the WADSTRATEGY and the LOW KNOWLEDGE group experiencedlower EFFORT when using the PROFILE STRATEGY.

This was followed by a complex interaction contrast toinvestigate the interaction effects of the independent variablesKNOWLEDGE and STRATEGY on the dependent variableEFFORT. The weighting for the complex interaction con-trast was consistent with the hypothesis being tested [48]. Acomplex interaction contrast to detect the interaction effectsof KNOWLEDGE (HIGH, LOW) and STRATEGY (com-bined EBA WAD, PROFILE) yielded a significant result

. From the above-mentionedseries of statistical analyses, it was concluded that there existsa strong interaction effect between the independent variables



KNOWLEDGE and STRATEGY on the dependent variableEFFORT. Subjects with HIGH KNOWLEDGE experiencesignificantly lower EFFORT when using the EBA STRATEGYor the WAD STRATEGY. Subjects with LOW KNOWLEDGEexperience significantly lower EFFORT when using the PRO-FILE STRATEGY.

VI. RATIONALE FOR EXPERIMENT 2

Given these differing responses to agent search strategiesbased upon the level of product class knowledge of the user,the next step was to examine more closely the specific char-acteristics of each strategy that contribute to this interaction,alterations to which could eliminate the interaction effect.

VII. THEORETICAL DEVELOPMENT OF

HYPOTHESES FOR EXPERIMENT 2

Decision environments that are characterized by highertradeoff difficulty and/or higher conflict will be associated withincreased negative emotion as well as increased tendencies fordecision makers to choose avoidant options [49]. The ability tochoose an avoidant option (skip making the decision) is foundto reduce this negative emotion [49]. Hence, the ability to avoidspecifying preference values for certain attributes or to providea neutral response (“don’t know” or “don’t care”) may reduceindividuals’ negative response toward the agent search strategy.Additionally, if individuals are permitted to provide a measureof their confidence in their preference specifications, they mayfeel more confident that the recommendation given by the agentreflects a weighting of the attributes in terms of their confidencelevel. It is expected that by modifying the application in thefollowing two ways, the negative affective reactions of subjectswith low product class knowledge toward agents/applicationswhich use the EBA and WAD strategies will be eliminated orsignificantly reduced.

1) Increasing the amount of information provided: When thesubjects are expressing their preference values for the at-tributes and when they are browsing the attribute infor-mation for the alternatives, instead of just being presentedwith the attribute names on the screen, they are presentedwith hypertext links for each attribute. If they click thehypertext link for any attribute, they are presented with ascreen of information that provides detailed informationabout that attribute.

2) Increasing the degree of control the subjects have in ex-pressing their preferences and making their decisions:This increase in the degree of control is achieved throughthree mechanisms:a) The subjects are given the option to skip attributes

when expressing preference or cutoff values for the at-tributes when they are using the WAD and EBA deci-sion strategies.

b) The subjects are given the option of specifying confi-dence levels in their preference specifications for eachattribute.

c) The subjects are given the option of returning to thepreference specification stage from any stage in thedecision process and changing the preference values

specified. Thus, after subjects have obtained some in-formation about attributes, they are able to utilize thatinformation in specifying their preferences.

Hypothesis 2: Changes to the functionality of the applica-tion such as: 1) increasing the amount of information providedin the description of each attribute and 2) increasing the degreeof control the subject has in expressing their preferences andmaking their decision, will reduce the negative affective reac-tions of subjects with low product class knowledge toward theagents/applications which employ the EBA and WAD strategies.

VIII. RESEARCH METHODOLOGY FOR EXPERIMENT 2

One hundred twenty MBA students participated in the study.The subjects used in Experiment 2 were different from the sub-jects used in Experiment 1. This was done to eliminate bias dueto history effects and learning effects. Measures of subjectiveknowledge for the product category were taken. Subjects iden-tified as belonging to the group “high product class knowledge”received further training in the product class. Subjects identi-fied as belonging to the group “low product class knowledge”did not receive further training. This was done to manipulateproduct class knowledge. Subjects in each of the groups “high”and “low” product class knowledge were then randomly as-signed to one of three treatment conditions—MODIFIED EBA,MODIFIED WAD, or PROFILE. This yielded a sample size of20 subjects for each of the cells in the 2 3 research design. TheHYPERTEXT strategy was omitted from Experiment 2 since itdid not yield any significant results in Experiment 1. Followingthe experiment, measures of affective response to each agentsearch strategy such as satisfaction with the decision process,confidence in the decision, propensity to purchase, perceivedcost savings, cognitive decision effort, and trust in the agent’srecommendations were taken.

IX. RESULTS FOR EXPERIMENT 2

Table VIII presents the cell means obtained for each ofthe dependent variables in Experiment 2 when the modifiedapplications were used as the treatment conditions. The 2 3,two-factor ANOVAs which were conducted did not yield anysignificant overall interaction effects between the independentvariables KNOWLEDGE and STRATEGY for any of the de-pendent variables SATISFACTION, CONFIDENCE, TRUST,PURCHASE, SAVINGS, or EFFORT. The planned contraststhat were conducted yielded only one significant result: thatsubjects in the HIGH KNOWLEDGE group expressed signifi-cantly higher positive affective preferences for the MODIFIEDEBA STRATEGY and the MODIFIED WAD STRATEGY ascompared to the PROFILE STRATEGY. Subjects in the LOWKNOWLEDGE group did not express significantly higherpositive affective preferences for the PROFILE STRATEGYas compared to the MODIFIED EBA STRATEGY or theMODIFIED WAD STRATEGY. Thus the interaction effects ofKNOWLEDGE and STRATEGY were significantly reduced bymodifying the EBA and WAD treatment conditions to increasethe degree of control provided to the user and to increase theamount of information provided to the user. Examination of thecell means shows clearly that the LOW KNOWLEDGE group



TABLE VIIICELL MEANS FOR EACH EMOTION AS A FUNCTION OF AGENT SEARCH

STRATEGY AND PRODUCT CLASS KNOWLEDGE FOR EXPERIMENT 2

responded much more positively to the modified EBA and WADstrategies than they did in Experiment 1.

X. CONCLUSION

The experimental results presented here demonstrate that thecognitive fit of agent search strategy and subjects’ product classknowledge has a significant impact on consumers’ ability to in-tegrate information, to understand inputs to their judgments,and to be confident about their judgments. This research hasimplications for a number of settings within the realm of mar-keting and consumer behavior, especially in the electronic com-merce environment. Understanding the reasons behind individ-uals’ varying reactions to agent strategies based on the level oftheir product class knowledge allows web retailers to modifytheir web sites and include functionality in their agents/appli-cations to make the agent’s search strategy more appealing toboth those with high as well as low product class knowledge.Specific recommendations to managers would include makingat least two agent search strategies (WAD or EBA, and PRO-FILE) available to users and allowing users to use the one theyprefer. Alternatively, managers could include the functionalitiesdemonstrated in Experiment 2 which reduce negative reactionsby individuals with differing levels of product class knowledge.

This research has demonstrated that optimizing the cogni-tive fit between agent search strategy and consumers’ productclass knowledge significantly increases consumer satisfactionwith the decision process. However, it has been noted that al-though some decision aids may improve decision making, abuseis possible [50]. The decision aid most likely to be a part of anelectronic commerce environment (i.e., a cutoff rule that allowsthe formation of a consideration set containing only those alter-natives that pass consumer-specified attribute cutoffs) can leadto suboptimal decisions in efficient choice sets [51]. A separatestream of research has shown that a second likely characteristicof an electronic commerce environment, i.e., a visually rich pre-sentation, can distort the decision process by diverting attention

away from information that is most important for the task athand [52], [53].

A. Managerial Implications

This research has significant managerial implications in termsof providing guidelines for the design of web sites in order tooptimize the interface between the consumer and the system.

1) Design of User Interfaces for Electronic Commerce: Thisresearch should help designers of web sites to make accurategeneralizations about the effects of computerized decision aidson strategy selection which will enable them to design theirweb sites so as to provide the optimal interface for a given taskenvironment.

2) Impact on Consumer Satisfaction and Confidence: Giventhe large number and variety of decision aids currently emergingon the Internet, it is imperative that we investigate how these de-cision aid formats impact consumer satisfaction and confidencein the decision. This is necessary to avoid consumer perceptionsof nonutility, and ultimately nonuse of the computerized deci-sion aids.

3) Marketing Strategy of Online Vendors: Many merchantswho have set up electronic shopping malls on the Internet fearthat the reduced cognitive search effort associated with this en-vironment will lead to increased price competition and lowerprofit margins [2], [54]. This may lead to merchants adoptinga strategy of providing a suboptimal web site so as to makeit difficult for consumers to use this medium to obtain pricecomparisons, quality comparisons and comparisons across websites. This research has demonstrated that optimizing the cogni-tive fit of agent search strategy with consumers’ product classknowledge results in an increased perception of cost savingsamong consumers and increased satisfaction with the decisionprocess. This will lead to the consumers using this channel moreextensively to search for prepurchase information and makingtheir purchases through this channel. Merchants who adopt thestrategy of not providing the optimal interface to consumers ontheir web sites will risk losing a substantial portion of the busi-ness that will be transacted via this channel.

B. Limitations of the Research

As with any experimental investigation, there are a numberof limitations present in this research. This research was re-stricted to the selection of cars on the Internet. Clearly, a va-riety of choice situations as well as products must be investi-gated before generalizable comments can be made to guide thedevelopment of computerized decision aids. Another limitationof this research is the composition of the group that partici-pated in the study. The sample is homogeneous, and presumablyhas greater than average cognitive capabilities. A more diversesample would have offered greater opportunity for generaliza-tion of the findings.

C. Directions for Future Research

This research has examined the influence of computerized de-cision aids on decision making and their impact on different as-pects of performance and satisfaction with the decision process.Much more work needs to be done on examining the influence



of these computerized decision aids on consumer preferences.Some of the potential research areas are discussed below.

1) Use of Query-Based Decision Aids in Different Informa-tion Environments: Ample evidence exists that information isnot simply acquired in reaction to pre-defined preferences, butthat it also helps decision makers define their own values andpreferences as they engage in the process of acquiring informa-tion [3], [55]. In other words, the information itself changes theway preferences are constructed, and therefore one cannot de-fine the decision space in advance. There exist potential liabil-ities for computerized decision aids in dynamic environmentsin which innovation can change the correlation structure of al-ternatives in the environment. As long as the information envi-ronment is stable and does not change much, the structure ofpreferences can be expected to have some stability. Hence insuch environments, computerized decision aids can be benefi-cial. However, in situations in which information changes overtime, consumers served by computerized decision aids alonewould be unlikely to notice the changing correlational structureof the environment. In these environments, additional mecha-nisms would need to be built into the system to continuouslyupdate the knowledge base. Expert systems could play a vitalrole in these decision environments.

2) Planning Upgrade Paths: Under some circumstances, itmight be better to have a simple user interface that does notrequire much effort to learn and use. The advantages of suchinterfaces are primarily at the initial stages of usage, whenknowledge is low. Over time, as knowledge accumulates, theadvantages of more powerful and flexible interfaces becomemore apparent. The challenge for marketing managers is toprovide consumers with information systems that change overtime such that they fulfill the consumers’ short term needswithout sacrificing the consumers’ long-term interests.

APPENDIX ASCALE ITEMS USED TO MEASURE THE CONSTRUCTS

The following scale items were used to measure the con-structs. The respondents were told to mark each of the responseson a seven-point Likert scale ranging from “Strongly Disagree”to “Strongly Agree.”

A. Independent Variables Being Manipulated

1) Product Class Knowledge (KNOWLEDGE):

K.1 I am an expert in cars.K.2 I am very experienced in purchasing cars.K.3 I am very knowledgeable about cars.K.4 I understand the features of cars well enough to eval-

uate the alternative car models.K.5 I am not at all familiar with cars (r).K.6 I have a great interest in cars.

B. Dependent Variables

1) Trust in the Agent’s Recommendations (TRUST):

T.1 I believe that the alternatives which the agent recom-mended to me were consistent with the preferences Iexpressed.

T.2 The agent can be trusted to recommend alternativeswhich closely match the preferences I expressed.

T.3 I am convinced that the agent recommended alterna-tives which most closely matched my preferences.

T.4 The alternatives recommended by the agent were notcredible (r).

T.5 The agent recommended alternatives which were con-sistent with my preferences.

T.6 The agent has probably used my preference specifica-tions in recommending alternatives to me.

T.7 It is questionable whether the agent used my prefer-ence specifications in recommending alternatives tome (r).

T.8 The agent can relied on to use my preference specifi-cations when it recommends alternatives to me.

T.9 The agent can be depended on to recommend alterna-tives which closely match my preferences.

2) Propensity to Purchase (PURCHASE):

P.1 I would like to purchase this car.P.2 If I purchased a car right now I would purchase this car

model.P.3 I would purchase this car if I had the money available.P.4 I feel a strong urge to purchase this car.P.5 I am willing to pay the price quoted for this car.P.6 It is very likely that I will purchase this car.P.7 I would definitely like to purchase this car.P.8 It is important that I purchase this car.

3) Satisfaction With the Decision Process (SATISFAC-TION):

S.1 This system is one of the best ways to select a car.S.2 If I could do it over again, I’d rather not use this system

to select a car (r).S.3 I am not happy that I used this system to select a car

(r)S.4 This system was very useful in helping me to select the

best car model to suit my requirements.S.5 If I had to select a car in future, and a system such as

this was available, I would be very likely to use it.S.6 If my friend was searching for information in order to

purchase a car, and I knew that a system such as thiswas available, I would be very likely to recommendthis system to him.

4) Confidence in the Decision (CONFIDENCE):

C.1 I am confident that I selected the best car model to suitmy needs.

C.2 I am confident that I selected the car model which bestmatches my preferences.

C.3 I am not confident that I selected the best car model(r).

C.4 There are probably other car models I should have ex-amined more closely (r).

C.5 I would select this same car model if I had to make thedecision again.

C.6 This is clearly the best car model available for mybudget.



5) Perceived Cost Savings (SAVINGS):

SAVE.1 By using this system to select a car, I was able toobtain the best value for my money.

SAVE.2 The use of this system has enabled me to save a lotof money in purchasing a car.

SAVE.3 If I had not used this system I would have obtaineda better deal for my money (r).

SAVE.4 This car is a real bargain.SAVE.5 This system enabled me to compare the prices of

different car models very efficiently.SAVE.6 I could have obtained a better deal from a car dealer

(r).

6) Cognitive Decision Effort (EFFORT):

E.1 The task of selecting a car model using this system wasvery frustrating.

E.2 I easily found the information I was looking for (r).E.3 The task of selecting a car model using this system took

too much time.E.4 The task of selecting a car model using this system was

easy (r).E.5 Selecting a car model using this system required too

much effort.E.6 The task of selecting a car model using this system was

too complex.

REFERENCES

[1] J. R. Bettman, E. J. Johnson, and J. W. Payne, “Consumer decisionmaking,” in Handbook of Consumer Behavior, T. S. Robertson and H.H. Kassarjian, Eds. Englewood Cliffs, NJ: Prentice-Hall, 1991, pp.50–84.

[2] J. W. Alba, J. Lynch, B. Weitz, C. Janiszewski, R. Lutz, A. Sawyer,and S. Wood, “Interactive home shopping: Incentives for consumers,retailers, & manufacturers to participate in electronic marketplaces,” J.Market., vol. 61, no. 3, pp. 38–53, 1997.

[3] J. W. Payne, J. R. Bettman, and E. J. Johnson, The Adaptive DecisionMaker, Cambridge, U.K.: Cambridge Univ. Press, 1993.

[4] N. Capon and M. Burke, “Individual, product class, and task-related fac-tors in consumer information processing,” J. Consum. Res., vol. 7, pp.314–326, Dec. 1980.

[5] J. Jacoby, R. W. Chestnut, and W. A. Fischer, “A behavioral approachin nondurable purchasing,” J. Market. Res., vol. 15, pp. 532–544, Nov.1978.

[6] W. L. Moore and D. R. Lehmann, “Individual differences in search be-havior for a nondurable,” J. Consum. Res., vol. 7, pp. 296–307, Dec.1980.

[7] W. G. Chase and H. Simon, “Perception in chess,” Cog. Psychol., vol. 4,pp. 55–81, Jan. 1973.

[8] , “The mind’s eye in chess,” in Visual Information Processing, W.G. Chase, Ed. New York: Academic, 1973.

[9] M. T. H. Chi, R. Glaser, and E. Rees, “Expertise in problem solving,”in Advances in the Psychology of Human Intelligence, E. Sternberg,Ed. Hillsdale, NJ: Lawrence Erlbaum, 1981.

[10] H. L. Chiesi, G. J. Spilich, and J. T. Voss, “Acquisition of domain-relatedinformation in relation to high and low domain knowledge,” J. Verb.Learn. Verb. Behav., vol. 18, pp. 257–273, Jun. 1979.

[11] J. H. Larkin, J. McDermott, D. P. Simon, and H. A. Simon, “Expert andnovice performance in solving physics problems,” Science, vol. 208, pp.1335–1342, 1980.

[12] R. D. Anderson, J. L. Engledow, and H. Becker, “Evaluating the relation-ships among attitude toward business, product satisfaction, experience,and search effort,” J. Market. Res., vol. 16, pp. 394–400, Aug. 1979.

[13] G. Katona and E. Mueller, “A study of purchase decisions,” in ConsumerBehavior: The Dynamics of Consumer Reaction, L. H. Clark, Ed. NewYork: New York Univ. Press, 1955, pp. 30–87.

[14] J. W. Newman and R. Staelin, “Multivariate analysis of differences inbuyer decision time,” J. Market. Res., vol. 8, pp. 192–198, May 1971.

[15] , “Prepurchase information seeking for new cars and major house-hold appliances,” J. Market. Res., vol. 9, pp. 249–271, Aug. 1972.

[16] J. E. Swan, “Experimental analysis of predecision information seeking,”J. Market. Res., vol. 6, pp. 192–197, May 1969.

[17] M. Sujan, “Consumer knowledge effects on evaluation strategies medi-ating consumer judgments,” J. Consum. Res., vol. 12, pp. 31–46, Jun.1985.

[18] A. Gershoff and P. West, “Using a community of knowledge to buildintelligent agents,” Market. Lett., vol. 9, no. 1, pp. 79–91, 1998.

[19] E. J. Johnson, “Expertise and decision under uncertainty: Performanceand process,” in The Nature of Expertise, M. T. H. Chi, R. Chi, and M.J. Farr, Eds. London, U.K.: Lawrence Erlbaum, 1980.

[20] J. Shanteau, “How much information does an expert use? Is it relevant?,”Acta. Psychol., vol. 81, pp. 75–86, 1992.

[21] M. Brucks, “The effects of product class knowledge on informationsearch behavior,” J. Consum. Res., vol. 12, pp. 1–16, Jun. 1985.

[22] J. Shanteau, “Psychological characteristics and strategies of expert de-cision makers,” Acta. Psychol., vol. 68, pp. 203–215, 1988.

[23] J. Mahajan, “The overconfidence effect in marketing management pre-dictions,” J. Market. Res., vol. 29, pp. 329–342, Aug. 1992.

[24] M. T. H. Chi, P. J. Feltovich, and R. Glaser, “Categorization and repre-sentation of physics problems by experts and novices,” Cog. Sci., vol. 5,pp. 121–152, Apr. 1981.

[25] M. T. H. Chi and R. Glaser, “The measurement of expertise,” in Ed-ucational Testing and Evaluation, E. L. Baker and E. S. Quellmalz,Eds. Newbury Park, CA: Sage, 1980.

[26] J. W. Alba and L. Hasher, “Is memory schematic?,” Psychol. Bull., vol.93, pp. 203–231, Mar. 1983.

[27] G. N. Punj and D. W. Stewart, “An interaction framework of consumerdecision making,” J. Consum. Res., vol. 10, pp. 181–196, Sep. 1983.

[28] G. F. Smith, “Toward a heuristic theory of problem structuring,”Manage. Sci., vol. 34, no. 12, pp. 1489–1506, 1988.

[29] J. Shanteau and T. R. Stewart, “Why study expert decision making?Some historical perspectives and comments,” Organizational Behaviorand Human Decision Processes, vol. 53, pp. 95–106, 1992.

[30] J. Shanteau, “Competence in experts: The role of task characteristics,”Org. Behav. Human Decision Process., vol. 53, pp. 252–266, 1992.

[31] J. F. Voss and T. A. Post, “On the solving of ill-structured problems,”in The Nature of Expertise, M. T. H. Chi, R. Glaser, and M. J. Farr,Eds. London, U.K.: Lawrence Erlbaum, 1988.

[32] A. Newell and H. A. Simon, Human Problem Solving. EnglewoodCliffs, NJ: Prentice-Hall, 1972.

[33] W. R. Reitman, Cognition and Thought: An Information Processing Ap-proach. New York: Wiley, 1965.

[34] H. A. Simon, “The structure of ill-structured problems,” Artif. Intell.,vol. 4, pp. 181–201, 1973.

[35] H. A. Simon, “Information processing theory of human problemsolving,” in Handbook of Learning and Cognitive Processes, W. K.Estes, Ed. Hillsdale, NJ: Lawrence Erlbaum, 1978, vol. 5, , pp.271–296.

[36] R. N. Taylor, “Nature of problem ill-structuredness: Implications forproblem formulation and solution,” Decision Sci., vol. 5, pp. 632–643,1974.

[37] J. Klayman, “On the how and why (not) of learning from outcomes,”in Human Judgment: The SJT View, B. Brehmer and C. R. B. Joyce,Eds. New York: Elsevier, 1988.

[38] E. J. Johnson and E. Russo, “Product familiarity and learning new infor-mation,” J. Consum. Res., vol. 11, pp. 542–550, Jun. 1984.

[39] J. N. Conover, “Familiarity and the structure of product knowledge,”in Advances in Consumer Research, A. Mitchell, Ed. Ann Arbor, MI:Assoc. Consumer Res., 1982, vol. 9, pp. 494–498.

[40] B. Walker, R. Celsi, and J. Olson, “Exploring the structural characteris-tics of consumers’ knowledge,” in Advances in Consumer Research, M.Wallendorf and P. Anderson, Eds. Provo, UT: Assoc. Consumer Res.,1987, vol. 14, pp. 17–21.

[41] J. W. Alba and J. W. Hutchinson, “Dimensions of consumer expertise,”J. Consum. Res., vol. 13, no. 4, pp. 411–454, 1987.

[42] S. Fiske, “Social cognition and social perception,” Annu. Rev. Psychol.,vol. 44, pp. 155–194, 1993.

[43] D. Maheswaran and B. Sternthal, “The effects of knowledge, motiva-tion, and type of message on ad processing and product judgments,” J.Consum. Res., vol. 17, pp. 66–73, Jun. 1990.

[44] P. Herr, F. Kardes, and J. Kim, “Effects of word-of-mouth and product-attribute information on persuasion: An accessibility-diagnosticity per-spective,” J. Consum. Res., vol. 17, pp. 454–462, Mar. 1991.

[45] S. J. Hoch and J. Deighton, “Managing what consumers learn from ex-perience,” J. Market., vol. 53, pp. 1–20, Apr. 1989.



[46] W. Wood and A. Eagly, “Stages in the analysis of persuasive messages:The role of causal attributions and message comprehension,” J. Person.Soc. Psychol., vol. 40, pp. 246–259, Feb. 1981.

[47] P. M. West, C. L. Brown, and S. J. Hoch, “Consumption vocabularyand preference formation,” J. Consum. Res., vol. 23, pp. 120–135, Sept.1996.

[48] R. Rosenthal and R. L. Rosnow, Contrast Analysis: Focused Compar-isons in the Analysis of Variance, Cambridge, U.K.: Cambridge Univ.Press, 1985.

[49] M. F. Luce, “Choosing to avoid: Coping with negatively emotion-ladenconsumer decisions,” J. Consum. Res., vol. 24, no. 4, pp. 409–433, 1998.

[50] P. Todd and I. Benbasat, “The influence of decision aids on choicestrategies: An experimental analysis of the role of cognitive effort,”Org. Behav. Human Decision Process., vol. 60, pp. 36–74, 1994.

[51] R. E. Widing and W. W. Talarzyk, “Electronic information systems forconsumers: An evaluation of computer-assisted formats in multiple de-cision environments,” J. Market. Res., vol. 30, pp. 125–141, 1993.

[52] S. L. Jarvenpaa, “The effect of task and graphical format congruence oninformation processing strategies and decision-making performance,”Manage. Sci., vol. 35, pp. 285–303, 1989.

[53] , “Graphical displays in decision making—The visual salience ef-fect,” J. Behav. Decision Making, vol. 3, pp. 247–262, 1990.

[54] Y. J. Bakos, “Reducing buyer search costs: Implications for electronicmarketplaces,” Manage. Sci., vol. 43, no. 12, pp. 1676–1692, 1997.

[55] A. Tversky, S. Sattath, and P. Slovic, “Contingent weighting in judgmentand choice,” Psychol. Rev., vol. 95, pp. 371–384, 1988.

Gene Rathnam is a Senior Partner at SatyamConsulting Services, Chicago, IL. He has publishedover 20 refereed journal articles and over 30 articlesin refereed conference proceedings. His researchfocuses on electronic commerce and informationsystems management. He is the recipient of severalgrants from the National Science Foundation.


556 ieee transactions on systems, man, and cyberneticsâ€”part

Documents