Spiekermann, Sarah: Online Information Search with Electronic Agents: Drivers, Impediments, and Privacy Issues

50

Chapter 4. Online Information Search for High Involvement Goods: A Structural Equation Modelling Approach

4.1 User Control in High-Involvement Online Searches with Agents

When consumers pursue targeted search online today, they do not chiefly rely on agents. Instead, most of the product search conducted on the Web is still done manually by visiting different Web sites, investigating product listings, descriptions or photographs. Some consumers do both, search manually and with the help of an agent when they shop online.

One main difference between manual information search and agent-based search is that in the former case the user “manually initiates and directly controls“ the information search process [Pazgal, 1999, p.1] while in the latter case he delegates a considerable part of the search responsibility to the autonomous software system. Thus, while the user-driven type of product search implies that consumers control the selection process from the total spectrum of available product offerings to a reduced consideration set, agent search implies that this act of selection is done automatically. As a result, the consumer loses control over a considerable part of the search process while at the same time saving effort due to task delegation. As was discussed above, this loss of control is a major challenge for agent acceptance and it is therefore regarded as a central aspect in human-agent interaction [Norman, 1994].

However, the discussion of user control reaches beyond concrete agent design issues. It is also at the centre of a debate on whether agents are at all sensible to use at the interface level [Shneiderman and Maes, 1997]. Shneiderman who is sceptical of using agents in the interface and instead proposes search that is directly manipulated by the user states: “The philosophical contrast is with ’user-control, responsibility, and accomplishment‘. Designers who emphasize a direct manipulation style believe that users have a strong desire to be in control and to gain mastery over the system...Historical evidence suggests that users seek comprehensive and predictable


51

systems and shy away from complex unpredictable behavior...“ [Shneiderman, 1997, p.36]. In contrast, agent proponents, even though they recognize the importance of user control and understanding, see a necessity in agent technology in order to “reduce work and information overload“ [Maes, 1994, p.1]. They argue that by ’making the user model available to the user‘ (e.g. with the help of comprehensive interfaces) sufficient control and understanding is achieved. Furthermore, they argue control may be sacrificed for other benefits such as time savings. As Maes states: “I don‘t mind giving up some control, actually, and giving up control over the details as long as the job is done in a more-or-less satisfactory way, and it saves me a lot of time“ [Shneiderman and Maes, 1997, p.54].

A key problem when it comes to agent-assisted targeted product search is that a consumer does not know for a long period of time whether the search process has been done in a ’more-or-less satisfactory way‘. Unlike for example an agent-based mail filtering system where a user can immediately check whether the agent has correctly sorted and filtered incoming messages, the quality and reliability of a consumer shopping agent can, to its full extent, only be judged upon at the moment the recommended product has been received or even tried out. Of course, it can be argued that the online consumer already regains search control once the agent has presented a consideration set, which he can then carefully examine before purchasing anything. Certainly, agent designers can also integrate control points into the systems; i.e. information on whether the agent was able to find independent product reviews or search reputation networks allowing users to decide whether the search has proceeded in a satisfactory way. However, as Widing and Talarzyk [1993] have pointed out, there is some risk that the selection of products made by agents could be sub-optimal for consumers<23> and there are certainly limits to what can be communicated to consumers for control purposes. As a result, consumers who actively search for products and explicitly order an agent to conduct part of the product search for them always have to trust the system to a certain extent, sacrificing some of their control. In future agent scenarios, where some academics envision software systems that take over the entire purchase process without referring back to the user [Borking et al., 1999; Pazgal, 1999], this problem of trust


52

and control will be even more serious. It is not surprising therefore that Urban et al. recognized in 1998 that the final and arguably most important requirement of a successful agent would be that it develops and maintains trust [Urban et al., 1998]. To the extend that consumers, however, do not develop trust in agents, they will probably continue to rely on manually controlled search forms or prefer directly manipulateable interfaces. The question is on what factors this trust in or reliance on an agent finally depends.

As was mentioned above, Urban et al.‘s [1999] study on the acceptance of a trust-based advisor agent produced some first insights into the conditions under which consumers are prepared to delegate tasks to agents. The degree of product knowledge a customer has on the product he seeks, his age and Internet experience were shown to be relevant drivers for agent use. Given this evidence, this thesis chapter raises the question whether there are not other user and context specific factors that can explain the degree of reliance on and information search with advisor agents.

For this purpose, user-controlled, manual search activity and agent-based search are investigated and compared as to their drivers and impediments. Gained insights are deducted into factors that explain some of consumers‘ wish for more or less control in a targeted information search process online.

The factors used to investigate targeted online information search are derived mainly from studies in offline information search behavior. These studies have discussed a myriad of potential drivers and impediments to impact and explain consumer search behavior (see [Beatty and Smith, 1987, p 86] for an overview). For example, it was found that besides product knowledge and experience [Punj and Staelin, 1983; Kiel and Layton, 1981] there are also factors such as perceived product risk [Sundaram and Taylor, 1998; Dowling and Staelin, 1994; Srinivasan and Ratchford, 1991; Capon and Burke, 1983] and purchase involvement [Sundaram and Taylor, 1998; Beatty and Smith, 1987] that determine depth and breadth of information search. Personal variables related to search have not been restricted to age [Katona and Mueller, 1955 cited by Beatty and Smith, 1987], but were also found in the form of attitude towards search [Thorelli and Thorelli 1997; Punj and Staelin, 1983], education [Claxton et al., 1974] or self-confidence [Kiel and Layton, 1981].


53

Investigating the influence of some of these traditional concepts on targeted online information search a structural equation model is presented below. The dependant variables in this model are manually controlled, user-driven information search and agent-based search. And as all potentially explanatory factors for behavior are tested separately for the two constructs, insights are being gained into what drives and impedes users to rely more or less strongly on an advisor agent instead of personally controlling information sources. Figure 5 gives an overview of the model tested. The next section reports on the concrete hypotheses integrated in it.

4.2 Model Constructs and Hypotheses

Abbildung 5: Model of Online Information Search: Unobserved Constructs and Stated
Directionality of Relationships


54

4.2.1 Endogenous Model Constructs

At the core of the structural equation model is the dependant construct of online information search. In line with the distinction between agent assisted and manually controlled search these two forms of online search are also distinguished for the current model. A number of drivers and impediments have been hypothesized to explain them.

One construct frequently investigated in the context of information search is perceived risk. Perceived product risk denotes a consumer‘s assessment of the consequences of making a purchase mistake, as well as of the probability of such a mistake occurring [Cunningham, 1967]. As a result of this initial risk assessment consumers were shown to engage in information search in order to reduce the perceived risk to an acceptable level. More precisely, risk was shown to be a multidimensional construct with consumers differentiating between functional, financial, social and psychological risk [Kaplan et al., 1974]. Functional risk is the uncertainty that a product may not perform as expected, financial risk that the product will not be worth the financial price and would have been available cheaper elsewhere, socio-psychological risk that a poor product choice will harm the consumer‘s ego or may result in embarrassment before one‘s friends, family or work group.

Probably, most risk dimensions relevant in the physical purchase process will continue to play a role in online environments. However, it could be that the degree to which individual risks are perceived is different online than offline. For example, as the online medium allows for much greater price transparency it may be that the financial risk of buying overpriced products is relatively low compared to little transparent offline markets. At the same time, being not able to touch and really see the product anymore, the socio-psychological risk might be higher in online markets than for their offline counterpart. In addition, there might be a new dimension of risk gaining relevance online, which is the delivery risk attached to a purchase. Buyers might fear that products won‘t arrive on time or be in perfect condition. Because


55

there was no delivery service included in the experimental store, delivery risk has not been included in the current model.

The influence of perceived purchase risk on information search has been investigated in a myriad of studies [Dowling and Staelin, 1994; Srinivasan and Ratchford, 1991; Kaplan et al., 1974; Cox, 1967]. Also for in-home shopping environments its relevance has been confirmed [Sundaram and Taylor, 1998]. In his meta-analysis of the risk construct Gemünden [1985] concludes, however, that perceived risk seems to be particularly valid for high-involvement goods and less so for commodities, because lower levels of product risk do not trigger information search as a risk reduction strategy. As a result of these findings, perceived risk has been included in our model of online information search. It was expected that higher levels of perceived risk would lead participants to use both means of search in a relatively intensive manner. As former models of information search have suggested a mediating role of risk between exogenous variables such as product knowledge and information search [Srinivasan and Ratchford, 1991], perceived risk was considered as an endogenous variable in our model and it was hypothesized that:

H1: The more product risk a consumer perceives prior to the purchase of a camera, the more he or she will interact with an electronic advisor agent.

H2: The more product risk a consumer perceives prior to the purchase of a camera, the more will he or she consult detailed product information.

4.2.2 Exogenous Model Constructs

Referring to earlier information search studies, the concepts of cost and benefit of search, product knowledge, product experience and purchase involvement were included in this model of online search.

A recognized construct in structural equation models of information search [Srinivasan and Ratchford, 1991; Punj and Staelin, 1983] (and theoretical reflections thereon) [Moorthy et al., 1997] are the costs and benefits of search. Costs of search in these studies represent the accumulation of physical and cognitive effort as well as


56

monetary expenditures necessary to find the right product. Benefits of search have been described as satisfaction with the product chosen or cost savings realized through the search activity [Punj and Staelin, 1983]. Benefits have also been recognized in relation to the degree of uncertainty present in the choice of environment, risk aversion and the importance a buyer gives to the product category sought [Moorthy et al., 1997].

In an online context, cost and benefits of search will probably continue to trigger search effort. Yet, especially the cost side may be of different nature online than offline. As was mentioned above, academics have pointed to a reduction of search costs in online environments [Alba et al., 1997]. In fact, many traditional search cost variables (such as the physical effort to travel to stores, the implied transportation cost or the cost of cognitive effort to handle the complexity of product comparison) may be comparatively less important in online environments than offline. At the same time, two traditional information search cost factors, namely information processing time and ease of access to information, were shown to continue to play a role for online environments, their design and consumer product choice [Lynch and Ariely, 2000; Hoque and Lohse, 1999]. Both of these cost factors are linked to the time investment a user is willing to make as part of the online search process.<24> Thus, even though the time required for an online search is already minimal in comparison with the offline world, it still appears to play a role in the way consumers search for information. As a result, time cost has been included in our model of online information search. While the named studies investigated the information search cost construct only for user driven information search, referring mostly to product listings, the model hypothesizes that time cost may be equally important in an interaction process with an agent. After all, consumers may weigh the number of specifications they make and potentially skip interactive search categories (in our case any of the 7-question cycles) in order to minimize time investment. Two hypotheses have been derived:


57

H3: The more time cost a consumer perceives while searching for product information, the less will he or she interact with an electronic sales agent.

H4: The more time cost a consumer perceives while searching for product information, the less they will consult detailed product information.

Costs of search have traditionally been outweighed by their benefits. For online environments this argument will probably continue to hold true. As in offline environments the benefits of search reside in the identification of an appropriate product. If consumers feel that interacting with an agent helps them to identify the right product they will probably be ready to invest into a relatively extensive dialogue (at least in a high-involvement context). If agent interaction is, however, thus beneficial, they will probably invest less effort into manual searching. To stress the relevance of perceived benefits from agent interaction for online information search, it was hypothesised that:

H5: The more benefits a consumer perceives from interacting with an agent, the more they will interact with it.

H6: The more benefits a consumer perceives from interacting with an agent, the less will he or she consult detailed product information.

Another construct that has continuously been shown to influence offline information search is product knowledge [Srinivasan and Ratchford, 1991, Beatty and Smith, 1987, Punj and Staelin, 1983]. Yet, what consumers actually know about a product category (objective knowledge) and what they think they know (subjective knowledge) is often differing and may have diverging effects on search [Brucks, 1985]. As a result, the empirical findings on how knowledge influences search have been contradictory. For the purpose of the current study there has been a focus on the knowledge consumers claim to have on a product category, because in the end it is this subjective feeling that will drive search effort. Subjective product knowledge was expected to limit searches by allowing responses to become routine or by allowing relevant information to be easier separated from the irrelevant, especially when interacting with an agent system. On the other hand, it was thought that higher


58

levels of subjective product class knowledge would lead subjects to increase manual search, since it allows one to delve deeper into information material. In addition, it was argued that those consumers who have more knowledge on a product also perceive less purchase risk [Sundaram and Taylor, 1998; Srinivasan and Ratchford, 1991]. It was therefore hypothesized that:

H7: The more knowledge a person states to have about a product category, the less will he or she interact with an electronic advisor agent.

H8: The more knowledge a person states to have about a product category, the more will he or she consult detailed product information.

H9: The more knowledge a person states to have about a product category, the less risk will he or she perceive when confronted with a buying situation in the respective category in an online context.

A concept that has gained considerable recognition in the study of information search and that has already been introduced in section 2.3.1. is the level of involvement a consumer has with the purchase situation (see e.g. [Beatty and Smith, 1987, Punj and Steward, 1983]). Involvement is seen as a motivational factor in consumer choice behavior and is attributed mainly to three causes [Deimel, 1989]: personal predisposition (i.e. subjective needs or goals), situational factors (e.g. time pressure) or stimulus-dependant factors (e.g. influence of product or communication). While situational involvement has been integrated in the model as a separate construct, stimulus-dependant involvement has been seized indirectly through the construct of product knowledge and perceived risk. Involvement was expected to play on both, agent interaction and manual search. A number of authors have argued that purchase involvement is also closely related to the consequences element of perceived risk [Beatty and Smith, 1987]. It was therefore hypothesised that:

H10: The more involvement a consumer has with a purchase situation, the more will he or she interact with an electronic sales agent.


59

H11: The more involvement a consumer has with a purchase situation, the more will he or she consult detailed product information.

H12: The more involvement a consumer has with a purchase situation, the more risk will he or she perceive when confronted with a buying situation in an online context.

A number of studies have addressed the subject of consumer interactivity, and information exchange with first generation computer mediated environments. For example, based on the theory of exchange developed in marketing literature, Swaminathan et al. [1999] tested the impact of vendor characteristics, transaction security, privacy concerns and customer characteristics on the likelihood of electronic exchange. Other studies observed the importance of secure financial transactions for consumers‘ perceived risk in online transactions [Parachiv and Zaharia, 2000]. By far the greatest research attention has, however, been attributed to the impact of privacy concerns on information exchange [Culnan and Milberg, 1999; Swaminathan et al., 1999; Hoffman and Novak, 1999] and to the existence of flow in online navigation [Hoffman und Novak 1996, 2000]. These two constructs, privacy and flow, have therefore been integrated in our online search model.

Privacy concerns of online users are a hotly debated issue. As mentioned above, studies confirm that consumers have great concerns about breaches of privacy [Pew Internet & American Life Project, 2000; Ackerman et al., 1999; Hoffman et al., 1999; Westin, 1996]. Ackermann et al. [1999], for example, found three distinct groups of online users with different levels of privacy concern: marginally concerned users, a pragmatic majority and privacy fundamentalists. Yet, despite these concerns many Internet users do not possess even rudimentary levels of online surveillance knowledge, and they do not use the available tools to protect themselves [Pew Internet & American Life Project, 2000]. As a result, the relationship between privacy concern and subsequent behavior is unclear. Would users restrict online exchange in order to protect themselves? Swaminathan et al. [1999] suggested in an empirical study among 428 users that this might be so. However, as is the case with most privacy surveys, they only based their model findings on questionnaire data and lag observations of consistent action. How might people react to a friendly


60

anthropomorphic agent that gives good product advice in exchange for private information?

Privacy can be sacrificed by both interacting with an agent, or by simply navigating online sites. All activities are usually logged by several servers that host the content displayed on users‘ screens. However, as was outlined above, when consumers interact with advisor agents on website (which ask for key-words or retrieve personal data through dialogue-based systems) they reveal a considerable amount of direct personal information. Consumers were therefore expected to be particularly cautious when using the interactive applications leading to the hypothesis:

H13: The more privacy concern a consumer expresses over the revelation of personal data, the less will he or she interact with an electronic sales agent.

Another phenomenon apparently occurring when navigating in online environments is ’flow‘. This flow status is a psychological state that has been investigated in the context of intrinsic motivation since the 1960‘s [Csikszentmihaly, 1995]. Hoffman and Novak observed its relevance for online environments [1996, 2000] and defined it here [2000, p.23] as a “state occurring during network navigation which is: (1) characterized by a seamless sequence of responses facilitated by machine interaction, (2) intrinsically enjoyable, (3) accompanied by a loss of self-consciousness, and (4) self-reinforcing.“ Thus, when consumers search for information online, it is possible that they lose perception of time and keep on navigating longer and in more directions than they initially planned to. Seen the creation of flow in online environments, the aim was to control this phenomenon with the following hypotheses:

H14: The more flow a consumer perceives, the more will he or she interact with an electronic sales agent.

H15: The more flow a consumer perceives, the more will he or she consult detailed product information.


61

Finally, it is intuitive to suggest that online consumers who used physical retail channels to get an overview of the product category and are thus more advanced in the buying process than their peers, would engage in less information search online than those who entered the online search process unprepared. The reason for this is that in interacting with an agent, informed customers might already know what selection criteria are the most important for them and are able not only to reduce the number of search criteria to a reasonably small size, but can also make up their mind more quickly regarding the specifications they prefer. As they know what they want, they may also be able to view product alternatives quicker and understand detailed product information more easily. Even though the stage in the buying process and product knowledge are related concepts, they have been distinguished for modelling purposes. Consumers could have felt knowledgeable about a product category without having gone to a store in advance of the online shopping trip. At the same time, subjects may have gone to a store before shopping online, but still felt little knowledgeable about the product category. Given this, it was hypothesized that:

H16: The further a consumer is advanced in the buying process, the less will he or she interact with an electronic sales agent.

H17: The further a consumer is advanced in the buying process, the less will he or she consult detailed product information.

H18: The further a consumer is advanced in the buying process, the less risk will he or she perceive when confronted with a buying situation in an online context.

4.3 Measures

4.3.1 Measurement of Endogenous Model Constructs

4.3.1.1 Measurement of the Information Search Construct

In the literature on offline information search, search activity has typically been operationalized by the time employed, the number of stores visited, the number of product alternatives inspected, the number of friends consulted etc. [Beatty and


62

Smith, 1987, Punj and Staelin, 1983]. For the purpose of the current study, measuring information search levels had to be adjusted to the electronic medium. While the relative amount of time spent searching was kept as one factor representing the search effort, the number of page requests was added as a second measure. Time was recorded for interaction with the electronic agent (phase 2) and for the two product inspection periods (phases 1 & 3). The time for interaction with the agent has been represented through the total time spent on answering agent questions and going back to the 7 category survey-page. The number of page requests in the context of agent interactivity stand for the intensity of exchange a user sought with the electronic agent. As was described above, the agent asked 56 purchase related questions, each of them representing a separate page. Users could return to this interactive functionality at any time during the shopping process and modify answers initially given. This activity of modifying specifications added to the number of pages requests in the interaction cycle as well as the time spent on the functionality. Finally, calls for the Top-10 ranking originating from the agent dialogue or the 7 category survey-page have been added to the number of page requests representing the breadth of agent interaction.

The number of individual product alternatives viewed added to the manual search construct. Each camera model on offer in the online shop was described on a separate html-page that could either be viewed in phase 1 or in phase 3. In addition to this detailed description, users had the possibility (in phase 3) to enlarge the photograph of each object in a separate page. The number of photo enlargements have been added as additional page requests to the construct of manual search. Finally, product descriptions were always requested from a page that listed the models available; either the Top-10 product ranking or the initial product orientation list (in phase 1). Together, product model lists, factual descriptions and photo enlargements made up the number of page requests for the dependant manual search construct. For all these pages time has been recorded and taken as a second measure. Both measures, time and page requests, are extremely precise measures of search when compared to the effort recall measures traditionally used in offline studies on information search.

Both time and page requests were recorded until a participant ended the search process which could be done either by pressing the ’buy-button‘ or the ’exit-button‘.


63

Time and page requests were also the only model constructs that were automatically recorded by the system. All the other measures were derived from participants‘ answers to pre- and post-shopping questionnaires. Appendix B1 gives again an overview of the different site pages and table C2 in Appendix C of the measures used.

It could be argued that the choice of time as a metric for the search undertaken is questionable since subjects have been asked to stay for a specified minimum of time at the lab. The time-cost factor that is usually present in shopping activities was therefore slightly manipulated. In fact, briefing the participants in this way may have led to a reduction in the variance of the time variable. However, the variance finally observed can be attributed more effectively to the constructs tested and is less subject to personal motivations in time management that would otherwise have gone uncontrolled. In addition, most of the subjects spent more time in the laboratory than they had to. It can therefore be argued that time is still a good measure; particularly as is was only important to observe the relative differences in behavior present in treatments with the same time conditions.

4.3.1.2 Measurement of Perceived Product Risk

Previous work was referred to in order to measure product category risk . As was outlined above, perceived risk has been characterized as a multidimensional construct with people differentiating between several negative consequences of a purchase including functional, financial, sociological and psychological risk [Kaplan et al., 1974]. For the current model, risk dimensions have been combined into one overall index (that has been proposed and tested by academics in earlier studies [Peter and Tarpey, 1975, p.30]). As a result, risk has been captured in the following way:


64

with ORPj = overall perceived risk for brand j

PL ij = probability of loss i from the purchase of brand j

IL ij = importance of loss i from purchase of brand j

n = risk facets (here n = 4)

OPR contains two components: “...a chance aspect where the focus is on probability [of losing] and a ’danger‘ aspect where the emphasis is on severity of negative consequences of purchase“ [Kogan and Wallach, 1964 cited in Peter and Tarpey, 1975, p.30]. Cunningham [1967] originally suggested a multiplicative relationship.

In the pre-shopping questionnaire, risk perception was measured by employing a 15-point scale for both dimensions, probability and importance of loss (see pre-shopping questionnaire in Appendix A5a). In order to calibrate the way in which different people respond to scales, each individual had to rate not only camera purchases, but also potential dangers and probabilities of loss associated with ’extreme products‘ in terms of risk, namely toothpaste and used automobiles.

4.3.2 Measurement of Exogenous Model Constructs

In order to measure time cost, earlier studies were considered which have introduced the idea of measuring time cost as opportunity cost. For example, Srinivasan and Ratchford [1991] measured time cost by asking people for their general time constraints and implied that this perception would be a measure for the opportunity cost perceived while searching for product information. In the present study, time cost was therefore grasped similarly by asking participants after shopping whether they had had the feeling during search that they would have rather done something else instead of sitting in a lab.

The problem in specifying the benefit construct is that, strictly speaking, benefits are not an antecedent, but a result of search. More precisely, perceived benefits of search


65

are the anticipated result of each additional search step performed [Moorthy, 1997; Weitzman, 1979]. Studies that measure the benefits of search should therefore try to capture either expected or ongoing benefits of search. This, however, has turned out to be a challenge. Either studies referred to the post satisfaction with the product bought [Srinivasan and Ratchford, 1991] or employed very general measures testing for consumers‘ backward belief in the merits of the search activity [Srinivasan and Ratchford, 1991]. Doing so, self justification may have impacted responses. On the other hand, measuring expected benefits of search prior to the actual search taking place carries the risk to prime subjects‘ behavior. The measurement problem was attempted to be circumvented by taking the perceived quality of agent recommendations as an indicator for perceived search benefits. Doing so, neither self-justification effects were present in our measure nor have subjects been primed. Instead, it has been possible to capture participants‘ ongoing impression of the quality of exchange, (closely linked to the identification of the right object.

For the measurement of product knowledge and involvement, measures have been used in the current study that have been proposed and tested in earlier empirical works [Srinivasan and Ratchford, 1991, Moore and Lehmann, 1980]. Table C2 in Appendix C gives a detailed overview of questions employed.

For the measurement of the two variables privacy and flow identified to be relevant for online environments parts of recent studies on these subjects have been employed. To measure privacy concerns some of the scales developed by Ackerman et al. [1999] were used. Participants were asked ten questions reflecting to what degree they would be ready to reveal certain types of information about themselves on a web site , including identification information (e.g. address or name) and profiling information (e.g. hobbies or income). The arithmetic mean of answers given to these 10 questions provided an index for participants‘ privacy concerns.

Flow is a construct that is relatively complex to measure. In psychological experiments conducted by Csikszentmihalyi et al. [1995], the so-called Experience Sampling Method (ESM) has been employed which involves permanent and unexpected measurement of the current state of consciousness during an activity. Thus, upon a notification signal of a transmitter that subjects have to carry with


66

them, they are required to respond to a short questionnaire (so called random activity information sheet) testing their current state of being. As a constant measurement of flow was not practicable in the shopping experiment, an additive index has been developed that is based on a number of questions capturing the flow experience as defined by Csikszentmihalyi et al. [1995] and Hoffman and Novak [2000]. The questions used to measure flow were derived from the random activity information sheets used in ESM experiments and attempted to capture what Hoffman and Novak [2000, p.24] characterized as the cognitive state of flow on the Web which would be “determined by (1) high levels of skill and control, (2) high levels of challenge and arousal, (3) focused attention and (4) is enhanced by interactivity and telepresence“

Finally, the fact that some participants had gone to a physical retail outlet was taken into account in advance of the experiment. There, some had already chosen products of interest for themselves that they now wished to buy for a 60% discount in our online store. Even though the online store made it difficult for them to rapidly identify their consideration set, because there were not brand names displayed, these subjects might still have behaved differently to those who were not informed. Subjects were therefore asked in advance of the buying session whether they had informed themselves of the product they wanted to purchase before coming to the lab and also to what degree they had already decided on products (consideration set). The two answers given were then combined to one index entitled Stage in the Buying Process.

Table C2 in Appendix C gives a detailed overview of all measures for the different constructs integrated in the equation model of information search. A major limitation of construct measurement is that constructs usually did not have more than 1 or 2 indicators. More precisely, the models captures 4 constructs (privacy concern, flow, perceived risk, stage in the buying process) with the help of an index, 4 other constructs (purchase involvement, product knowledge and the online search variables) with the help of 2 indicators and finally, costs and benefits of search with only one indicator. The reason why model constructs had to be concentrated in this way is that for equation modelling the recommended ratio of sample size to number of free parameters is about 5:1 [Bentler and Chou, 1987]. As was mentioned above, the study was restricted in terms of sample size, which implied that the number of


67

free model parameters had to be minimized. Using reliable indices as construct representatives was a reasonable strategy to do so.

4.4 Results

4.4.1 Data

Before model estimation, the data (see table 2) was screened for outliers<25> which led to an exclusion of 6 from a total of 151<26> observations. In addition, 29 subjects had missing data, which could have been imputed [Little and Rubin, 1987]. However, imputing missing values by using a Maximum-Likelihood approach implies a multivariate normality assumption. As this assumption does not hold true for our data basis<27>, model estimation had to be based on 116 cases.

4.4.2 Model Estimation and Fit

A structural equation modelling approach was used to simultaneously test model constructs and their relations. This approach was chosen, because it allowed for the test of complex relationships between constructs and also, to some extend, operationalized theoretical constructs by multiple items. The model was estimated by the software program Mplus [Muthén and Muthén 1998] which uses Maximum-Likelihood Method (MLM) as a standard modelling approach. Yet, since data


68

deviated from the normality assumption that underlies a Maximum-Likelihood (ML) estimation it was necessary to use the more robust MLM estimation option available in Mplus. This MLM estimation approach respects the condition of a relatively small number of observations as well as deviations from normality distribution. It usually has an effect on estimated standard errors for parameter estimates as well as the Chi-square test statistic.

In an initial model estimation thus conducted with MLM, adequate fit indices were obtained. However, four of the latent variable indicators had negative measurement error variances. These so-called “Heywood cases“ are a problem often encountered in structural equation modelling under the conditions of a small sample size and only two indicators per latent variable [Boomsma 1982; Anderson and Gerbing 1984]. As neither sample size nor the number of indicators could, however, be changed, the problem of improperty was solved by employing a strategy pursued by earlier studies on information search where negative error variances have been set to zero [Punj and Staelin, 1983]. Recalculating the model with the time variable for manual search being set to zero resolved the negative error variance problem for the entire data set. In addition, modification indices that can be generated by ML-estimation suggested a considerable increase in model fit by specifying a covariance between the measurement errors of two search indicators, namely the number of page requests during the interaction with the agent as well as those requested for manual search. From a theoretical point of view this correlation makes, in fact, sense in that the two constructs of agent interaction both represent facets of information search for which some unobserved but common variable carries explanatory value.

Standard fit measures in structural equation modelling obtained for the final model are highly satisfactory (see table 3) [Homburg and Baumgartner, 1995]. The RMSEA is considerably below the cut-off value of .05 [Browne and Cudeck 1993; Hu and Bentler 1999] and both CFI and TLI are above the threshold value of .95 [Hu and Bentler 1998]. <28> Table C2 in Appendix C contains the system output corresponding to the results reported.


69

Tabelle 3: Fit Measueres for Model of Online Information Search:

Overall Model Fit

(MLM)

RMSEA = .038

CFI = .974

TLI = .952

The rather small sample size prevented a highly sophisticated operationalization of the theoretical constructs by multiple indicators. Nevertheless, based on parameter estimates for the model, the reliability and validity of our two-indicator measurement models has been assessed (see table 4). For this purpose indicator reliability was used [Bagozzi, 1982], factor reliability (squared correlation between a construct and an unweighed composite of its indicators; see [Bagozzi and Baumgartner, 1994]) and the average variance extracted [Fornell and Larcker, 1981]. Both, factor reliability and average variance extracted can be regarded as measures for convergent validity. Since all these values were above the required threshold values [Bagozzi and Yi, 1988] and as factor loadings were all significant, the construct measurements can be regarded as reliable and valid (see table 4).


70

Tabelle 4: Reliability and Validity of Measurement Models:

Factor

Indicator

Indicator Reliability

Factor Reliability

Average variance extracted

Involvement

1

2

.908

.431

.841

.747

Product class knowledge

1

2

.978

.438

.811

.688

Interaction with agent

1

2

.848

.455

.761

.615

*Product inspection

1

2

1.000*

.626

.864

.761

Required level

ge.4 ge.6 ge.5

*error variance fixed to zero

4.4.3 Model Relationships

Fit measures of the model indicate that the overall relationships hypothecated to exist for online information search sufficiently reflect reality. Interesting for the better comprehension of online information search is, however, to what extend the hypotheses made hold true and at what level of significance they can be supported. Figure 6 gives an overview of the findings (for detailed output data see Appendix C, table C3).


71

Abbildung 6: Antecedent Variables and Directionality of Relationships for a Model of Online Information Search

In hypotheses 1 and 2 it was postulated that the more purchase risk a consumer perceives the more will he or she search for information. In fact, hypothesis 1 that users use an electronic agent more intensively when they perceive higher levels of risk was not confirmed by the data. In contrast, it was observed that participants tended to rely less heavily on the interactive functionality the more risk they perceived, even though this relation is not significant. At the same time, they consulted significantly more detailed product information the more risk they perceived, confirming hypothesis 2. This finding suggests that consumers may engage more in manually controlled forms of search the more product risk they perceive. At the same time, they do not necessarily like to rely on an interactive functionality like agent Luci. In the section 4.5. below this phenomenon is commented on in more detail.

All exogenous constructs that were hypothesized to influence the perception of risk, namely product knowledge (H9), purchase involvement (H12) and the stage in the


72

buying process (H18) proved to be in the right direction. However, none of them were statistically significant, except for product knowledge.

As far as the time cost of search is concerned, hypothesis 4 was supported. The data revealed that the more participants had wished to do something else while shopping online, the less they manually sought for information. The same was true for agent interaction (hypothesis 3), however not to a significant level. The results might indicate that agent functionality is relatively less impacted by consumers‘ time constraints than are user-driven search forms. This, however, would have to be proven by more research.

In contrast to hypothesis 5, the more benefits a user derived from their interaction, the less he or she was willing to invest in the interaction process. In fact, since that benefits of search were measured in the form of perceived accuracy of agent recommendation, it is intuitive to argue that the better the initial recommendation made by the agent, the less participants had an incentive to return to the interactive functionality to enhance or modify search parameters. However, even if this explanation is straight forward, the finding is still interesting because it raises awareness that one of the most basic assumptions made in information economics, which is that the more benefits one retrieves from information search, the more one searches for information, might be significantly impacted by agent technology (at least if benefits are measured in terms of identifying the right model). This impact resides in the possibility that the perceived utility of search renders decreasing marginal returns of search much quicker than this was the case for offline markets. The result is an inverse relationship between perceived search benefits and the activity of search. More research is certainly needed to investigate this finding and test its impact on the cost-benefit construct in information search theory. Hypothesis 6 that the more benefits a consumer perceives from interacting with an agent, the less will he or she consult detailed product information was supported by the data, however not at a significant level.

The traditional concept of product knowledge proved to be a reliable indicator for the prediction of interaction with the agent. Hypothesis 7 that the more knowledge a person states to have about a product category, the less will he or she interact with an


73

electronic sales agent was shown to be significant at the highest level. Also, the positive effect of product knowledge on manual search was in line with the initial hypothesis (H8), though not at a significant level. Thus, people who think that they know a lot about a product relied less on an advisor agent, spending less time and effort on interaction with it. At the same time, they had a slight tendency to invest themselves more in manual search.

Another traditional search factor which proved highly significant for both parameters of search, agent interaction and detailed product inspection, was product involvement (H10 and H11). The more involvement a participant had with the purchase situation, the more he or she used both information sources available from the online store.

In summary, most of the traditional information search factors identified for offline markets were supported by the online model, with more than half of them at a significant level. Only two relationships did not hold true, namely the impact of perceived risk, and search benefits on the interaction process with the agent.

Hypothesis 13 that privacy concerns would be negatively related to consumer willingness to interact with the agent system was confirmed by model results. In fact, the data does not only support hypothesis 13, but also suggests that privacy concerns may have the strongest impact on agent interaction amongst all variables tested. This finding means that marketers who employ highly interactive technologies on their web sites should, in their own interests, pay attention to the privacy conditions they offer to their customers. However, it should also be noted here that in average more than 85% of the agent‘s questions were answered by the participants. This is surprising, because answering agent questions is much more informative about a user than his navigating a site. Users‘ privacy concerns seem to have expressed themselves more in a restriction of navigation (measurable in time and page requests) than in a reduction on information disclosed. Seeing the contradiction of these findings and also the relevance of privacy for the Net community, privacy preferences and behavior are investigated in more detail in chapter 6 of this thesis.

The flow construct introduced by Hoffman and Novak [1996, 2000] for Web navigation proved significant to the model. The data confirmed that participants who


74

perceived more flow searched significantly more manually (hypothesis 15). This positive effect was, however, not significant in as far as the shopping agent was concerned (hypothesis 14).

Finally, the data supported at a non-significant level that the more participants were advanced in the buying process, the less would they interact with the advisor agent (hypothesis 16) or manually search for information (hypothesis 17). As there were no brand names displayed in the store, the strength of this finding must, however, be regarded with caution. In case of brand display the negative effect on information search could have been stronger, with participants going directly for their consideration set.

4.5 Discussion: Strategies of Information Search With or Without Agents

An interesting finding of the structural equation model was that both higher levels of perceived product risk and product knowledge did not seem to lead to higher levels of interaction with the agent.

The more product knowledge a participant stated to have about cameras, the less he interacted with agent Luci. At the same time a positive relationship was observed regarding manual search. This goes in line with Urban et al.‘s findings [1999], who found similar evidence that subjects with higher levels of product knowledge reported to prefer less reliance on an advisor-agent. Does this mean that consumers generally appreciate agents less the more they know about a product category? In order to investigate this question, the data was analysed in more detail.

The results of the structural equation model as well as those obtained by Urban et al. [1999] were impacted by the type of agent employed in the experiments and its specific perception by users. Both systems offered an in-depth dialogue system and wished to support a cross-section of product knowledge levels. As a result, some highly knowledgeable customers may not have found the level of expert-exchange they wished for. In short, reduced levels of interaction (actual or reported) could also


75

be attributable to low satisfaction levels with the very agent system employed in the experiment.

In order to investigate this argument, the relationship between subjective product knowledge and the level of satisfaction with the advisor agent was analysed which was measured after the shopping session. First, the two questions that had been employed to measure subjective product knowledge (KA, KB) were correlated with satisfaction levels (SL).<29> A negative correlation would suggest that more knowledgeable customers did not appreciate then interaction with agent Luci which indicates that the specific agent employed in the current experiment was not ideal for more knowledgeable customers. In case of a positive correlation, support would be given to the argument that, even if more knowledgeable users appreciated the agent system, they were generally less relying on it for their product choice.

Table 5 indicates a negative correlation between product knowledge and satisfaction with agent Luci. The more knowledge a participant stated to have in comparison to the average citizen (KA), the less did he appreciate the agent which is expressed in a significant negative correlation coefficient CORR (SL, KA ) = -.167*.

The correlations suggest that lower levels of interaction could be attributable to the failure of the very agent system employed in the experiment to serve the needs of highly knowledgeable customers. As a result, it cannot be argued that, in general, higher levels of product knowledge lead to lower levels of interaction with agents. More research on this aspect would certainly be of interest.


76

Tabelle 5: Relationship between Subjective Product Knowledge and Satisfaction with the Search Engine :

A:Subjective Level of Product
Knowledge (KA):

In comparison to the average citizen I
already know quite a lot about hifi-
equipment (e.g. stereos, cameras, TVs..)

5 = very true

4 = quite true

3 = depends

2 = not really

1 = not at all

B:Subjective Level of Product
Knowledge (KB):

I regularly advise peers in the choice of
their electronics.
 

5 = very true

4 = quite true

3 = depends

2 = not really

1 = not at all

CORR (SL, KA ) = -.167* (*p = .044)

CORR (SL, KB) = -.016 (p = .848)

Interpreting model results on perceived risk, similar observations were made as to the use of the two search forms offered in the online store: The more risk participants perceived prior to a purchase, the less they relied upon agent interaction (non significant relation) and the more they searched manually for information on each


77

object. As argued above, this finding suggests that consumers may rely more on manually controlled forms of information search the more risk they seek to reduce.

Again, in order to support this type of generalized argument, it was important to exclude the possibility that it was the quality of exchange offered by agent Luci in particular that led to the observation of the relationship. Investigating the relationship between perceived purchase risk prior to shopping (RP) and general satisfaction with the shopping agent (SL), however, suggests that satisfaction with the search engine and risk perception are two relatively independent constructs in our data; the correlation coefficient CORR (SL, RP) = - .069 being small and not significant. Also, when looking into the relationship between risk and satisfaction with the agent recommendation quality (SR) this independence is maintained displaying a non-significant correlation coefficient of CORR (SL, RP) = -.060.<30> Thus, the observation that participants used the agent in a relatively restricted manner the more risk they perceived cannot be attributed to low levels of satisfaction with the system or its recommendations (for SPSS output file see Appendix D, table D2). In addition, and as was outlined above, the agent dialogue was explicitly designed to address all major dimensions of risk with 64% of questions addressing functional, 9% financial, 9% sociological and 18% psychological risk. The experimental data therefore suggest that the more purchase risk a participant perceived the less he chose to rely on the automatic recommendation technology, seeking instead the control over the choice process. Of course, more research would be needed to confirm this finding which may be an indicator for the degree of acceptance (or reluctance) agent technologies will face when being deployed in high-involvement and high-risk electronic commerce environments.

4.6 Conclusion

The structural equation model proposed for drivers and impediments of online information search displayed a very good level of fit and supported the majority of hypotheses made. As a result, it was possible to show that determinants of information search identified in offline studies, including product knowledge,


78

purchase involvement and time cost, seem to hold true for the online world. Furthermore, (prove could be made of the influence of new variables such as privacy concerns and the achievement of a flow status for information search behavior in electronic environments.

As far as agent based versus manually controlled forms of search are concerned, it is interesting to see that consumers who perceived higher levels of risk prior to the purchase relied less strongly on agent advice than their peers and preferred to control the search process manually through product inspection. As far as product knowledge is concerned, the data suggest a similar tendency for more knowledgeable customers to rely less on agent advice. However, more research would be needed to confirm this finding. In addition to these potential impediments for agent use, risk and product knowledge, it was interesting to see that perceived time cost led to a smaller influence on agent interaction than on manually controlled forms of search. At the same time, agent interaction seemed to create less flow.

In line with the hypothesis made on privacy, expressed privacy concerns of participants seem to have led to reduced levels of interaction with the agent. However, this is a curious finding, since participants answered in average over 85% of agent questions. Thus, decreased levels of interaction stand in sharp contrast to the actual information disclosed. A more detailed analysis of this behavioral phenomenon is presented in chapter 6.

All in all, valuable insights have been gained on drivers and impediments for online information search with advisor agents and/or manually. An important limitation of the structural equation model presented though is the limited sample size on which it is founded. Also, the fact that there were only a few indicators per construct, often only one index, must be regarded as a drawback. On the other hand, the advantage of the structural equation modelling approach was that one could capture relationships simultaneously and avoid problems of multi-co-linearity often present in regression analysis. The good model fit supports our approach. If there had been serious problems in model set-up, the equation model would not have converged.


79

Finally, the model was built on only one product category. As other scholars in information search have pointed out, this single-measure, single-product variety limits though the generalizability of the findings [Beatty and Smith, 1987]. As a result, it would be interesting to challenge the findings on a bigger sample size and across several product categories. This was unfortunately not in the scope of the current thesis. However, one finding of particular interest that resulted from the structural equation model was still investigated for another product group: the influence of perceived risk and uncertainty on agent use. Do consumers really seek for more controlled information environments when they shop for higher risk products? Do they tend to rely relatively less on agent advice when they perceive more risk? This question was investigated in more detail, by comparing the concrete search activities participants displayed for winter jackets with those for compact cameras. The next chapter (5) reports on the results of this analysis.


Fußnoten:

<23>

Widing and Talarzyk [1993] showed that using attribute cut-offs to screen alternatives tends to result in inferior product decisions due to inadvertent product elimination.

<24>

In the Hoque and Lohse study [1999] information processing effort was, in fact, measured on the basis of time investment per online task (such as moving the mouse) while in the study conducted by Lynch and Ariely [2000] this time investment was implied through the experimental set-up.

<25>

The respect of time measures for information search in the model required an outlier analysis in order to take account of two phenomena: 1) some users had proceeded to the first page of ’orientation‘ without the experimenters‘ consent and before having answered the prior-to-shopping questionnaire. Even though orientation was interrupted, time was recorded for these participants on the respective level. 2) some users had to leave for the restroom during the shopping session.

<26>

As summarized in table 1, the original data set included 152 subjects. However, 1 subject did not answer the correct questionnaire version prior to shopping and therefore had to be excluded from analysis.

<27>

Using PRELIS 2.30 [Jöreskog and Sörbom, 1996] the assumption was tested that the variables are normally distributed. The multivariate tests (see for example [Bollen, 1989]) after listwise deletion of 29 cases with missing data show that the remaining data is, however, significantly skewed (z = 5.42, p = .000) while multivariate kurtosis represents a borderline case (z = 2.45, p = .014). An omnibus test on multivariate skewness and kurtosis (chi2 = 35.37, p = .000) further indicates that the data is not normally distributed, although deviation from the norm seems to be rather modest and in the first place concerns indicators for information search behavior.

<28>

To further support model validity the MLM-fit was challenged by additionally re-calculating the model with the more standard Maximum Likelihood (ML)-estimation. Here a moderate fit was confirmed with an RMSEA = .077, CFI = .919 and TLI = .851. However, since for small samples these fit criteria tend to over-reject true population models [Bentler and Yuan, 1999; Hu and Bentler 1998] these values should be regarded with caution.

<29>

Satisfaction with the agent (SL ) was measured by asking users after the shopping session: “What level of comfort did you perceive in interacting with the search engine?“ Participants answered on a 14-point scale from 1= no comfort at all to 14 = very high level of comfort

<30>

Recommendation quality (SR) was measured by asking participants after the shopping session on a 5-point scale: “How well did you perceive product recommendations to fit your needs?“


© Die inhaltliche Zusammenstellung und Aufmachung dieser Publikation sowie die elektronische Verarbeitung sind urheberrechtlich geschützt. Jede Verwertung, die nicht ausdrücklich vom Urheberrechtsgesetz zugelassen ist, bedarf der vorherigen Zustimmung. Das gilt insbesondere für die Vervielfältigung, die Bearbeitung und Einspeicherung und Verarbeitung in elektronische Systeme.

DiML DTD Version 2.0
Zertifizierter Dokumentenserver
der Humboldt-Universität zu Berlin
HTML - Version erstellt am:
Mon Dec 23 13:23:15 2002