User’s preferences in comparison shopping websites

(1)

User’s preferences in comparison

shopping websites

A Choice-Based Conjoint analysis on Shopbot

By

Nghi Hue Luong

# s2496917

University of Groningen

Faculty of Economics and Business

MSc Marketing Intelligence

Supervisor: Dr. M.C Non 2nd Supervisor: Dr. H. Risselada

(2)

i

Executive summary

Purpose – This study aims to examine and understand consumers’ preferences for a shopbot. By

identifying the relative importance and utilities of five different attributes, the findings can assist those involved in designing and managing shopbots to improve and optimize the use of shopbots based on the principal of effort reduction.

Design/methodology/approach – Choice-based conjoint analysis (CBC) was used to obtain

responses from a sample of 202 people. Participants were asked to choose a preferred shopbot from three hypothetical shopbots differing on five attributes. Two questionnaires were randomly distributed were each contained six choice sets and one hold out set. An OLS regression was further performed to assess the validity of the CBC outcome.

Findings – The results suggest that on aggregate level, users place the highest value on low

waiting time and price information displayed in a shopbot. Five segments were identified; “Information Oriented Users”, “Efficient Users”, “Discount Oriented Users”, “Optimizers” and “Simplicity Oriented Users”, in which each segment placed highest value on respectively price information, waiting time, discount bundling, number of alternatives and type. Demographic variables failed to significantly classify the respondents into segments as opposed to the personality trait; maximizer and satisficer. Differences in personality trait with respect to

preferences for price information displayed and bundling option were found between maximizers and satisficers. Validation of the CBC with ordinary least square regression confirmed the effect of personality traits in determining the importance of discount bundle as well as waiting time.

Keywords: Shopbot, choice based conjoint, preferences, search criteria, waiting time, number of

(3)

ii

Acknowledgments

This thesis marks the end of a journey as a master student as well as the completion of a double degree program in MSc Marketing Intelligence and MSc Strategic Marketing Management from respectively University of Groningen and the cooperating university, BI Norwegian Business School. The process of writing the thesis has been both challenging and rewarding. There are several people I wish to recognize who have either directly assisted or indirectly inspired the creation and completion of this master thesis endeavor.

First of all, I want to especially thank my supervisor, Mariëlle Non, for her helpful guidance and constructive feedbacks. Your advice has helped me to keep a thoughtful perspective during this long and sometimes arduous journey and your help and kindness is very much appreciated. Also, I want to thank my second supervisor for his help and suggestions as well as his teaching during the writing of the thesis, which has given me inspirations along the way.

Secondly, I am very thankful for the master thesis group members and a special thanks to my thesis partner for all the great conversations, ideas sharing and coffee breaks. It made the process of writing the thesis much more pleasant.

Thirdly, the thesis could not have been completed without a sufficient number of respondents. Therefore, I need to thank all my respondents for taking their time to fill in my survey and providing me the necessary data.

Furthermore, I would like to express my greatest gratitude to my family and friends for their support and encouragement when times were tough and things were not so perfect. A special gratitude is directed to my parents who have been so patient and given me the opportunity to study abroad. Your continued support has been heartwarming and motivational.

(4)

iii

List of figures and tables

Figure 1: Conceptual Model ... 14

Figure 2: Random choice set task with three options... 17

Figure 3: Fixed hold out choice task with three options ... 17

Figure 4: Percentage people with the corresponding personality trait score ... 22

Figure 5: Aggregated part-worth utilities for ordered values ... 24

Figure 6: Relative important attributes for each class ... 27

Figure 7: Relative Importance for "Information Oriented users" ... 30

Figure 8: Relative Importance for "Efficient users" ... 31

Figure 9: Relative Importance for "Discount bundling oriented users" ... 32

Figure 10: Relative Importance for "Optimizers" ... 33

Figure 11: Relative Importance for "Simplicity oriented users" ... 34

Table 1: Top 10 shopbots indicated by Google 13th of May 2012 ... 2

Table 2: Attribute and levels ... 18

Table 3: Demographic characteristics of the participants in the CBC analysis ... 22

Table 4: Attribute preference rating ... 23

Table 5: Utility parameters and relative importance of the aggregate model ... 23

Table 6: Model fit aggregate model ... 24

Table 7: Fit Criterion of 1 to 5 class model... 25

Table 8: Comparing 2 class model with 5 class model ... 26

Table 9: Utilities and relative importance of each segment ... 28

Table 10: Hit-rate across 2 models ... 29

Table 11: Hypotheses results... 35

Table 12: OLS results across 5 models ... 36

Table 13: OLS assumption test results ... 37

(5)

iv

1.0 Introduction

With today’s technological advances the World Wide Web is becoming more and more the main source for people to find information, purchase products and services, keep up to date on the latest news, trends and other such life events. An article in Washington Post posted by Ylan Q Muli in December 2007 stated that shoppers spend 175 billion dollars online every year – and it is predicted to grow by 20 percentages or more annually. The opportunity with internet and e-commerce is proliferative and it is the future. Furthermore, Nielsen NetRatings (2007) shows that the ability to efficiently compare offers is one of the most popular reasons for consumers to shop on the Internet. Many consumers buy online because they like the convenience or the price (Chiang and Dholakia 2003).It is argued that the internet reduces buyers search costs by

providing easy information retrieval (Pereira 2005). However, with the increasing growth in online business, consumers have more choices of web stores to shop, which challenges consumers’ ability to choose vendors to shop from. For consumers with a specific item to purchase in mind, it will be time-consuming to search through all different vendors who can possibly provide the item with an appropriate price and service quality. Furthermore, Johnson et al. (2000) found that consumers only search on one or two online stores for a purchase even though hundreds of competing web stores are just “a click away”. The increasing urge for efficient purchase and consumers’ interest to find the “best deal” has naturally led to a growth of popularity in using price comparison recommendation agents, or shopbots which is the

shortening for shopping robots. Shopbots are automated tools that allow customers to easily search for prices and product characteristics from a large number of online vendors. Due to this, consumers can potentially reduce the cognitive effort associated with decision making. By using various sources this means that one shopbot will give the information of ten or more sites with one simple search. Not only does this allow for good deals, but it also spares consumers the hassle of going to each and every one of those ten websites, saving time and energy.

Shopbots on the Word World Web today

A number of shopbots have emerged since the first shopbot, BargainFinder, was introduced into the world of e-commerce in 1995. According to Montgomery et al. (2004), adoption of shopbots showed a substantial increase from 0,1% in June 1997 to 5,7% in May 2002. Over time,

(8)

2

attracted some attention in academic research in recent years (Wan 2009; Allen and Wu 2010). Like most product development in competitive markets, shopbots should be developed to match user search preferences –if such preferences are known (Montgomery et al. 2004). Recent studies have examined the impact of shopbots on online consumer search (Zhang and Jing 2011) and electronic markets (Smith 2002). However, the preferences for a shopbot are not known. Because the number of comparison websites has grown large in numbers and areas, it is of the researcher’s interest to understand what shopbots are being valued the most. This study focuses on consumer’s preferences for a shopbot when searching for flight tickets. The main reasons for the chosen focus in the context of flight ticket comparison are as following. First of all, the travel industry has been ranked as the top online transaction category1, which in many ways reflects the future purchase behavior of flight tickets. Secondly, the number of current shopbots provided today are broad and proliferative (e.g. Expedia.com, farecompare.com, skyscanner.net), which makes it an interesting area of research. For instance, a search on the keywords: “flight ticket

comparison” on www.webstatschecker.com indicated up to hundred of online shopbot sites that provided price comparison on flight tickets. Table 1 illustrates the top 10 shopbots or domains with the strongest position in Google indicated 13th of May 2012. Thirdly, to the best of my knowledge, no previous research has used flight tickets as an example in the context of shopbots.

Table 1: Top 10 shopbots indicated by Google 13th of May 2012

The large scope of shopbots in comparing flight tickets put pressures not only on airline companies, but also among shopbot sites. Moreover, the vast amount of choices available can overwhelm consumers particularly under limited time and may cause them to make “wrong” decisions. Which shopbot would a consumer choose among hundreds of them? What criteria or preferences are consumers looking for in a shopbot to feel satisfied and use it again? As Punj and Moore (2007) stated, being able to consider a variety of options and being able to do so quickly are often valued the most. However, the desire to accomplish both goals is known to be the

(9)

3

“paradox of choice” in a Web-based store (Tamaki 2005). How do consumers actually reconcile the desire for finding the best option according to their needs with the objective of saving time and effort? How do consumers make trade-offs in selecting a shopbot?

1.1 Problem Statement and the Aim of the Study

Just as with almost any marketing instrument, shopbot web sites in general are not of the “one size fits all” category. The aim of this thesis is to examine the relative importance of shopbot attributes and to identify potential segments with respect to consumer’s user preferences. In reality, when people choose among various options of shopbots to use, they are faced with choosing between combinations of attributes. In the case of a shopbot of flight tickets for example, people would make the decision to use a particular shopbot by considering multiple factors such as waiting time, quality of information and recommendation provided, and search criteria options related to the use of the shopbot site. Drawing on the economics and marketing literature, and building on the research studies of Montgomery et al. (2004), who address the problem of designing a better shopbot, this master thesis will aim to provide insight into

consumer preferences for an optimal shopbot site. Furthermore, the paper will take into account the study of Punj and Moore (2007), who use the cognitive cost model to address which type of shopbot is preferred, as well as respond to Drechsler and Natter (2011), who suggest future directions to investigate the effect of price charts in a service context.

The research question for this thesis is:

RQ: How to optimize a shopbot according to consumer’s preference? What determines a

person’s decision of choosing a particular shopbot and can consumers be segmented accordingly?

1.2 Sub-questions

In order to answer the research question more thoroughly several sub questions are developed. By looking at the several shopbots available online (e.g., skyskanner.net, kayak.com,

travelsupermarket.com), it shows that there are differences between shopbots in terms of price information displayed, number and quality of alternative suggested, how long one have to wait until a suggestion appears, and if it offers bundle discounts. Taken into account these difference features of a shopbot the following sub-questions are answered:

(1) What is the utility of search criteria in choosing/using a shopbot? (2) What is the utility of waiting time in choosing/using a shopbot?

(10)

4 (4) What is the utility a consumer attaches to the price information displayed on the shopbot?

(5) What is the utility a consumer attaches to the bundling option on the shopbot? (6) Are there differences between segments in preference toward a shopbot?

Viewing consumer’s preferences as a whole may be too myopic as markets can be distinguished by number of smaller homogeneous segments in response to differing preferences (Wedel and Kamakura 2002). The last sub question will therefore take into account heterogeneity between users of a shopbot.

1.3 Contribution on the marketing literature and practice

Prior studies on shopbots have only focused on consumer goods such as books, computer hardware (Zhang and Jing 2011; Montgomery et al. 2004) , and apartments (Punj and Moore 2009, Punj and Moore 2007). To the best of my knowledge, no research has yet studied how to optimize a shopbot site in the context of a service. Therefore, this thesis will broaden the scope and shed light on the optimal preferences of a shopbot based on a flight ticket service context. Additionally, it examines whether the shopbot preferences differ between segments. Identifying distinctive groups based on the different preferences of a shopbot gives the company insight into how to design a shopbot to best target the segments. Little is known about the relative

importance of a range of elements or features in shopbots on people’s choice of behavior. Given the diversity of shopbot designs and its attributes, it is relevant to ask what attributes or features yield the best utility in determining the use of a specific shopbot.Acknowledging that shopbots earn revenue not only from choice, but also from consideration (e.g., shopbots can be paid if a visitor clicks on a link to the retailer site), it is important to understand how profitability of a shopbot can be optimized. The findings will be useful for marketers and shopbot operators who want to improve or implement a shopbot that is more accessible, inviting and valuable for users as well as increases traffic and thereby boosts the revenue. Furthermore, by understanding the consumers preferences and trade-offs, marketing managers can target their market more efficiently.

(11)

5

2.0 Literature review, hypotheses and conceptualization

2.1 Categorization of shopbots

Smith (2002) categorized shopbot design into three generations: stand-alone, contextual, and personalized. Stand-alone shopbots are the simplest form, and only provide consumers

comparisons across vendors’ offerings, as opposed to providing information about the product itself. Thus, stand-alone shopbots are useful for consumers who already have determined a specific product to buy and do not need additional product quality information. Stand-alone shopbots can be unbiased or biased. In early 2001, most shopbots migrated to biased listings where vendors can pay a fee for priority positioning in the shopbot’s comparison tables (Smith 2002). This change acknowledged vendors the importance to stay competitive and enhanced the revenue options available to shopbots. Second generation shopbots provide price comparison with additional product information in a contextual setting. Furthermore, the third generation of shopbots personalizes and compares differentiated products on the basis of the preferences of individual consumers. The Value Shopper service is one example of a personalized shopbot. For instance, if the consumer were to buy a camera, he or she would enter the desired zoom, size, price sensitivity levels and functions as preferred. Then, the system would return available products sorted based on the criteria put forward by the consumer. Likewise, many shopbots in the context of finding cheap flight tickets are personalized. Consumers first have to enter desired destination, date of departure/arrival, and possibly price range in order to get alternative offers which match the initial selection criteria.

(12)

6

2.2 Attributes

2.2.1 Search criteria – “smart” versus “knowledgeable”

The distinction between “smart” and “knowledgeable” shopbots has been previously noted in the literature (Maes 1999; Russo 1987; Punj and Moore 2007). A “smart” shopbot has the same information filtration capability as the “knowledgeable” one, but goes a step further by also suggesting alternatives that nearly or almost meet the selection criteria. In many instances, the initial selection criteria used by the consumer in a shopbot setting correspond to more a “wish list” of features sought than a realistic assessment of what may be available in the marketplace. Imagine there were no flight tickets that correspond to the “wish list” fulfilling the initial criteria. Here, the feedback mechanism in a “smart” shopbot will benefit the customer in greater effort reduction (Maes 1994) by suggesting flight tickets that closely meet the selection criteria.

“Knowledgeable” shopbots on the other hand, have no feedback mechanism and would therefore not give any alternatives if the initial criteria were not met. Thus, the “knowledgeable” shopbots focus on exactly matching the criteria of the consumer. The results from the study of Punj and Moore (2007) suggest that use of a “smart” shopbot enables consumers to maintain a focus on effort reduction because the next closest option will be suggested, whereas “knowledgeable” shopbots focus on providing the consumer better product “fit” or suggestions that only meet the initial criteria, and therefore less focus on effort reduction. It is worth mentioning that only when “no exact matches” of wished and offered alternatives are found does a difference between smart and knowledgeable shopbots arise. There are no differences in the feedback provided or

alternatives offered by the two types of shopbots (smart and knowledgeable) when “exact matches” were found.

With the aforementioned review of the literature, and taking into account the third generation of personalizing shopbots and the benefit of the feedback mechanism of a “smart” shopbot, it is believed that a shopbot providing search criteria with “smart” feedback mechanism will be more valued by the customers in comparison to the use of search criteria with “knowledgeable” recommendation. Therefore, the following hypothesis is:

H1: “smart” search criteria will give a higher utility than “knowledgeable” search criteria.

2.2.2 Waiting time

(13)

7

Montgomery et al. (2004), where cost includes the time to use the shopbot, the time to wait for the shopbot to respond, and additional cognitive effort. Waiting time or the response time is the time associated with retrieving the set of offers (Montgomery et al. 2004). Depending on the immediate needs of the consumer, waiting time may or may not be consequential. However, in general, the longer the waiting time, the higher the cost to the consumer (Talaga and Tucci 2001). In the case of a shopbot comparing flight tickets, it is reasonable to believe that the amount of waiting time desired by consumers is relatively small, because the main purpose of using a shopbot is to save time and because the volatility of price can change drastically in the travel/flight industry2. For instance, if you are only few seconds too late with booking a ticket, you may risk having to pay a higher price. Reducing search time and waiting time will all act to reduce some of the “extended'' or non-monetary costs of acquisition of goods and services (Talaga and Tucci 2001).

On the other hand, longer waiting time may be perceived positively as an indication of quality assurance. Thus, the consumer may find the shopbot more reliable and worthwhile, and therefore would be willing to trade the waiting time with better provided personalized offers. However, there are strong indications of disutility to waiting from previous studies of Internet behavior (Montgomery et al. 2004, Talaga and Tucci 2001). Moreover, a study of usability research shows that consumers place high emphasis on download times (Udo and Marquis 2001). Furthermore, Konana et al. (2000) conjecture that there is a direct trade-off between waiting time and costs. Dellaert and Kahn (1999) show experimentally that waiting can negatively affect evaluations of websites. Their results also suggest that waiting is not purely a function of time but can be mediated by other factors such as the size of the web application (e.g. size and number of large graphics and files of a web page), the infrastructure of the internet (for complete review see Rose & Straub 2001), the technological configuration of the web server and/or web user. These factors are however considered as out of scope in the current study and will not be elaborated further. Therefore, based on the aforementioned studies and the context of reducing cost and enhancing usability, it is believed that waiting time is negatively associated with what consumers prefer of a shopbot. Hence, the following hypothesis is:

H2: Waiting time has a negative effect on the level of utility of the consumer.

2

(14)

8

2.2.3 Number of alternatives

In similar vein with waiting time, number of alternatives or offers displayed on the shopbot may also be seen as cost, as greater information means increased cognitive effort required to process and compare alternatives (Steckel et al. 2005; Montgomery et al. 2004). On the other hand, more alternatives may increase the propensity of finding good deals. Moreover, the study of Allen and Wu (2010) showed that shopbot size (how many vendors it covers) is positively associated to shopbots’ market representativeness. Therefore, from one point of view, having access to a large number of alternatives is highly desirable (Johnson and Payne 1985). At the same time, the limited cognitive resources may simply unable the consumer to process the potentially vast amount of information about these alternatives (Häubl and Trifts 2000).

The notion of cognitive overload has a long history in consumer behavior research (Jacoby et al. 1974; Jacoby 1984; Keller and Staelin 1987). One challenge with current shopbot design is the number of alternatives presented, as every additional alternative presented will force the user to expend cognitive effort and will result in information overload. The increase in shoppers’ cognitive efforts may decrease the popularity of using a particular shopbot (Basartan 2001). Furthermore, previous research has shown that consumers are willing to trade off cognitive effort in the decision-making process for accuracy (Johnson and Payne 1985).Moreover, as the amount of information increased, Jacoby et al. (1974) asserted that consumers would feel better even though they actually made poorer purchase decisions. The opposite results were found in Malhotra (1982) and Lee and Lee (2004). Despite the diverse results of the effect information overload has on consumers there is no doubt of the existence of trade-off between effort and accuracy (Chen, Shang and Kao 2009).

Based on the aforementioned literature review and the notion of information overload, it is reasonable to believe that fewer numbers of alternatives reduce the cognitive burden. On the other hand, larger number of alternatives increases the likelihood to make a quality decision, but considering many alternatives takes time and effort. Therefore, this study hypothesizes that:

H3: The utility for a consumer is highest for a moderate number of alternatives

2.2.4 Price information

(15)

9

doubt that price is a crucial element to include in the shopbot site as it is a price comparison agent that aims to search for products at best prices from major online retailers (Harrington & Leahey 2006; Smith 2002). Previous studies on shopbots have assumed that consumers are searching for a perfectly homogenous product, comparing only prices, and will buy from the retailer that offers the lowest price they can find (Chen & Sudhir 2004; Iyer & Pazgal 2003; Smith 2002). Smith and Brynjolfsson’s (2001) results suggest that customers are more sensitive to shipping price and sales taxes than they are to the price of the item (base price). Findings from Morwitz, Greenleaf and Johnson (1998) say the opposite is true when it is difficult to associate the base price with surcharge (e.g., shipping price and taxes). In the context of flight tickets comparison, the display of the total price is most common. Other surcharges such as taxes (which shall be included in the total price) or additional luggage charge will mainly be displayed only in the further process in the selection process. Also, shipping cost is not applicable in the context of flight tickets as most tickets are being electronically sent or printed out from the self-service machine at the airport.

Price chart information and consumer purchase decisions – the impact of visualization Presenting price charts is a rather new phenomenon and has recently been found popular and applied in shopbots for consumer products. Besides providing distributions of actual prices, shopbots such as NexTag.com, SkinFlint.co.uk, and PriceScan.com have recently introduced line charts displaying a product’s full price history (Drechsler and Natter 2011). Despite of this increasing trend, to the best of my knowledge, there is very little discussed in the marketing literature on the effect price history visualized in a line chart and chart pattern characteristics have on consumer preference and decisions. Based on the study of Drechsler and Natter (2011), which examined the price charts’ influence on consumers’ price expectations and purchase intention on consumer goods that normally experience significant price decrease over time, this research builds on their work by examining another setting in which the supply of the product category is limited (e.g., air travel industry) and where prices fluctuate, decrease and increase over time (e.g., flight tickets).

Studies in behavioral finance show that the visualization of past prices itself leads to a change in investment decisions (Mussweiler and Schneller 2003; Brealey, Myers and Allen 2007).

(16)

10

tables (Vessey 1991). Whereas line charts especially facilitate the recall of trends, the tabulated representation of data on the contrary, facilitates the recall of specific amounts (Washburn 1927; Vessey 1991). Therefore, whenever it is important for consumers to quickly and easily recognize characteristics of data such as trends, volatility and functional relations, visual representation and line charts are preferred.

From a shopbot’s perspective, the provision of a price chart is worthwhile because it is perceived as containing valuable information from a consumer’s and a policy maker’s perspective which serves as an anchor to support consumer purchase decisions. Furthermore, according to Drechler and Natter (2011) it increases the popularity of shopbots. Since visualization may help in

reducing the information overload (Darlin 2006), one may ask if displaying a price chart will have a positive impact on consumers’ preference of using a specific shopbot over another. To the best of my knowledge, little is known of what type of price information yields better consumer and user preference. Therefore, to test for the utility attached to price information of past price history using a price chart, it is reasonable to include price line chart only, total base price only (flight ticket price) and combination of both the total price and line chart presented, to allow for comparison. It is believed from a subjective (exploratory) perspective that providing only line chart of past price history may be too simplistic and does not reveal the true cost information in comparison to an explicit representation of the total base price of a service. However, the combination of both, which provides explicit and visualized price information that may add additional value without increasing information overload, is expected to have a positive effect on utility. As such, the following hypothesis developed is:

H4: The combination of total price and price chart displayed on shopbots yields the highest utility for consumers.

2.2.5 Bundling

(17)

11

offersbundle promotion, and at the same time eases consumers’ cognitive effort and provides shoppers more effective and value-added services – if such service is preferred.

The study by Garfinkel et al. (2006) is the first research incorporating bundle pricing and promotional deals into shopbot design. Even though their findings show that number of items in the bundle suggests potential benefits, such as increased savings, caution must be made to generalize the findings in the real shopbot environment due the small sample size and time constraints of the study.

The use of bundle pricing and promotions has been a common marketing practice for a long time, and the increased use may be due to retailer’s avoidance of direct price competition pressure caused by shopbots. Deterministic and non-deterministic are among others the various bundling strategies that have been implemented by retailers (Simon and Wuebker 1999).

(18)

12

consumer may simply ignore the possibility of the cross discount bundling without extra effort. Hence, the study hypothesizes that:

H5: “Cross-discount bundling” yields a higher utility than no bundling in consumers’ choice preference of a shopbot.

2.3 Moderators

While the direct effect of the attributes mentioned above are important to understand consumer’s preferred choice and the trade-off between attributes, it has been argued that it may be somewhat obvious and that it is more meaningful to investigate moderating effects of external factors such as consumer and situational factors (Ranaweera, McDougall and Bansal 2005; Ajzen 1991; Baron and Kenny 1986). The paper will further examine the moderating effect of personality traits categorized by Schwartz et al. (2004) on consumer’s choice preference of using a specific shopbots and the utility attached to the choice.

2.3.1 Personality trait: Maximizer and Satisficer

Although historically, the science of economics has relied on the rational choice theory which assumes that people are rational choosers (von Neumann and Morgenstern 1944), modern behavioral economics has acknowledged that the assumption of complete information that characterizes rational choice theory is implausible (Schwartz et al. 2002). Given the wealth of information and the increase in product choices available in the online marketplace, Schwartz (2004) suggests that the increase in options has shifted accountability of making the best product choice from the firm to the consumer. Based on a maximization scale which measures individual differences in maximizing tendencies, Schwartz classified people into two overarching

categories; Maximizers and Satisficers. A Maximizer tends to engage in more product

comparisons, takes longer to decide on a purchase and analyzes extensively to seek out and find the elusive best option. Furthermore, studies have reported that maximizers are more likely to experience regret after a purchase, and feel less positive about purchasing decisions (Schwartz 2004, Nenkov et al. 2008). On the other hand, a Satisficer tends to search until encountering an option that crosses the threshold of acceptability, also referring to “good enough” option in comparison to “best option” in the case of maximizing.

(19)

13

maximizers and satisficers in the context of shopbots specifically. It remains to be determined whether maximizers also consistently act differently (or prefer different levels of attributes) than satisficers in the context of a shopbot. This paper responds to the future research direction proposed by Schwartz et al. (2002), which called for more understanding of the differences between maximizers and satisficers when it comes to actual choice behavior. In this context, that is, when choosing or using a shopbot, do maximizers prefer to examine more options before purchasing, suggesting more alternatives and information are preferred? Do they prefer a smart search criteria type of shopbot in comparison to satisficers? It is a common assumption that people tend to maximize their utility when making a choice. Based on the above discussion of the personal characteristics, it would be reasonable to believe that a Maximizer would prefer a smart shopbot which displays the combination of total price and price chart, more numbers of alternatives offered, less waiting time and a shopbot of cross-discount bundling option, in comparison to Satisficers who are more looking for the easy “good enough” options. The following hypothesis derived is:

H6: The more a consumer is a maximizer the higher utility it attaches to (a) smart search

criteria (b) less waiting time (c) higher number of alternatives (d) combination of total price and price chart displayed and (e) bundling option.

2.3.2 Demographic characteristics

The effects of customer demographics are likely to be context dependent and have a moderating impact on customer’s choice in choosing a specific shopbot (Ranaweera et al. 2005). While acknowledging the potential moderating effect of personality traits, other demographic characteristics such as age, gender, education level, occupation, and nationality are also

examined as moderator variables for comparison purposes, particularly as a means of profiling the expected differences between segments. There are no a priori propositions developed for each characteristic. However, there is a general agreement that demographic characteristics play an important role in marketing since they provide the opportunity to customize products and services, in this case shopbots, to better meet consumer needs (Ranaweera et al. 2005).

2.4 Conceptual framework

(20)

14

the five attributes; type of search criteria, waiting time, number of alternatives, price information and cross-discount bundling option. The researcher expects that the combination of different levels of attributes will yield differences in utility depending on personality traits such as if one is a maximizer or satisfier, and demographic characteristics such as age, gender, education level, occupation and nationality. These moderating variables are also called covariates.

Figure 1: Conceptual Model

Waiting time Type Number of alternatives Price information Cross-discount bundling Utility (Preference) H1 H2 (-) H3 H1 H4 H5 (+) Personality Trait Age, Gender, Occupation,

(21)

15

3.0 Methodology

This chapter of the paper describes the methods used in the study detailing the research strategy and design, as well as the questionnaire design with emphasis on the part of CBC experimental design. Furthermore, data collection and sampling will be described.

3.1 Research strategy

The current paper takes on a positivism approach which emphasizes descriptive quantitative research. There are specifically two types of analysis in focus; mainly Choice Based Conjoint analysis to understand the customer choice preference of a shopbot, and an OLS regression analysis as a validity check of the respondent’s choices. As previous studies (e.g. Moore 2004) have indicated, the results of attributes importance may differ based on the way questions are being asked. For instance, the CBC experimental part asks the respondent indirectly with a trade-off approach, whereas the OLS part asks and analyzes the importance respondents find in each attribute directly without considering other attributes. The strategy to employ both

methodologies enables the researcher to draw unique perspectives from each (Huber et al. 1993), and to gain confidence in the results. Prior to the CBC analysis, a rotated factor analysis is performed to identify and verify the 6 item Maximizer Scale constructs suggested by Nenkov et al. (2008), as well as to reduce the number of parameters in the CBC analysis to obtain a

parsimonious model.

3.2 Questionnaire design – a flowerpot approach

(22)

16

important for them on a 7-point Likert scale, where 1=strongly disagree and 7= strongly agree. These questions were used to conduct an ordinary least square regression analysis for validation purpose of the CBC outcomes and as a method to test the moderators, which will be elaborated later in the section.

3.2.1 Conjoint Experiment Design.

The third and main part of the survey contains the choice based conjoint design which takes on a full profile format randomly generated by the program Sawtooth SSI web 6.6.6. Based on a design efficiency greater than 95% (Appendix 3), 2 versions of the questionnaire were created, each containing 6 randomized choice sets and 1 fixed hold out set. By creating 2 versions or questionnaires, the aim is to maximize the accuracy of the estimated measures and avoid

overloading respondents with too many questions. Furthermore, because each choice set contains 3 profiles, each respondents evaluated in total 21(7*3) profiles including the fixed choice set, which exceeds the suggested minimum number of 9 (13-5+1) profiles that must be evaluated by each respondent to gain statistical efficiency and reliability of the results (Hair et al. 2010, 280). The profile displayed contains a level from every attribute in the study, and because the

Complete Enumeration method is used to generate the profiles, they are “almost-but-not-quite” orthogonal. This means that the attributes levels are duplicated as little as possible or have “minimal overlap” in generating profiles within choice sets. Within respondents, the profiles are nearly orthogonal and each frequency of level combinations between attributes is equally

balanced. This method is chosen as it is argued to be more efficient overall than purely

orthogonal design, particularly in this research where attributes have different numbers of levels (asymmetric design) (CBC v6.0 Technical Paper). In this case, only one profile occurred twice, however in different questionnaire versions.

Each choice task includes 3 profiles and respondents were asked to choose the one shopbot they prefer the most. Figure 2 shows an example of a choice set in the choice-base conjoint (CBC) questionnaire. Note that for the attribute type, the term “knowledgeable” shopbot used in the literature is replaced by “simple” with the aim to reduce confusion and make it easier for

(23)

17 Figure 2: Random choice set task with three options

The one hold out task which involved a fixed scenario used in both versions was used to assess external validity (i.e., the actual preferences of respondents on the fixed hold out task are compared with those predicted by the CBC analysis). Using the individual utility coefficients and their corresponding probabilities, the hit rate will be calculated. As such, the predictive performance of the hold out sample can be compared to the internal sample to check for model accuracy. The hold out set was designed by the researcher based on the consideration of

matching each alternative with the expected preference by each expected group (maximizer and satisficer) and for the purpose of segmentation with latent class analysis. Figure 3 illustrates the fixed hold out set which was placed as the 4th question in the CBC design.

Figure 3: Fixed hold out choice task with three options

(24)

18

3.2.1 Attributes and attribute levels

The CBC study limits the number of attributes to the following five attributes; search criteria, waiting time, number of alternatives, price information and bundling, based on a review of the literature and with the purpose of creating actionable and realistic alternative product concepts for shopbots with regards to flight tickets. Table 2 shows the different attributes and levels selected with the associated literatures. Regarding type of search criteria, two levels were chosen reflecting the “smart” and “simple” search criteria outcome of shopbots. Concerning the next attribute, waiting time, three levels were chosen based on a small experimental test prior to the study; 20, 50, and 80 seconds. The third attribute, number of alternatives, contains three levels based on observed alternatives provided in today’s shopbots (e.g. skyscanner.net,

farecompare.com, kayak.com etc.). The three levels are; 10, 50 and 100 alternatives. For the fourth attribute, price information, three levels were selected based on traditional and new modern ways of presenting information of price. These three levels are; price line chart only, total base price only, and combination of total price and line chart. Finally, bundling was chosen to have two levels to reflect the option of no bundling and 50% cross-bundling discount option.

Table 2: Attribute and levels

Attributes Levels Literature

Type of search criteria

 Smart  Simple

Maes (1994;1999); Russo (1987); Punj and Moore (2007)

Waiting time

 20 seconds  50 seconds  80 seconds

Montgomery et al. (2004); Talaga and Tucci (2001); Konana et al. (2000); Dallaert and Kahn (1999)

Number of alternatives

 10  50  100

Montgomery et al. (2004); Steckel et al. (2005); Häubl and Trifts (2000); Jakoby(1984)

Price information

 Price line chart only  Total base price only  Combination of base price

and price chart

Smith (2002); Zhang & Jing (2011), Pathak (2010); Drechsler & Natter (2011)

Bundling _ No bundling 50% cross-discount

Garfinkel et al. 2006

The concept of estimating choice based conjoint models, also known as discrete choice

modeling, lies in the concept of utility maximization with the use of multinomial logit or MNL approach (Louviere et al. 1983). With the selected attributes and levels known, the utility function of the choice for each segment j is determined as follows:

(25)

19 p(p=1,2), respectively for attribute X1=type; X2=Waiting time; X3=Number of alternatives; X4=Price information and X5= Cross-discount bundling.

As stated by Pindyck and Rubinfeld (2005), the utility is a numerical score that represents the satisfaction that a consumer gets from a given product. The higher the utility, the more satisfaction a consumer or in this case a segment will get out of the shopbot. Thus, the probability of choosing the shopbot increases with higher utility.

3.2.2 Personality trait as moderator.

Part four of the questionnaire contained 6 statements which sought to identify the respondents’ personality traits related to maximizer and satisficer. The 6 items extracted from the study of Nenkov et al. (2008) were measured on a 7-point Likert scale. The higher (lower) the score, the more likely the respondents are identified as maximizer (satisficer). Refinements of the questions were made due to outdated statement. Particularly, a statement such as “renting video is really difficult” was modified to “if I want to watch a movie at home, I am always struggling to pick the best one”. The modification of the question is believed to be more appropriate and relevant in todays’ situation.

3.2.3 Demographics

Part five asked the respondents to fill in simple demographic variables such as gender, age, education level, occupation and nationality which were all measured on nominal scale except for age and education, which were scaled and ordinal respectively. The reasons underlying the placement of demographic characteristics at the end of the questionnaire is because (1) it is not the stated primary objective of the research, but provides useful additional data to add a “face” or profile each segments, (2) it lies in the notion of unwillingness that stems from the fact that most respondents do not understand their relevancy to the study’s main information objectives and view them as inappropriate (Hair, Bush and Ortinau 2006). Finally, the survey concluded with a final thank-you statement.

3.3 Data Collection and Sampling

(26)

20

73% have used a shopbot to compare products or services before, and actually bought a flight ticket through a shopbot respectively. This indicates that a large portion of the sample has some sort of familiarity and understanding with the concept of a shopbot.

3.4 Data analysis 3.4.1 Factor analysis

For the purpose of the choice based conjoint analysis, the 6-items personality trait scale was factored into one component for the sake of simplicity and to avoid multicollinearity.3 Both KMO and Bartlett’s test of Sphericity implies that factor analysis is appropriate and the Cronbach alpha was 0.631. Although the generally agreed recommended level of Cronbach alpha is 0.7, the reliability between 0.6 and 0.7 also may be acceptable provided other indicators of a model’s construct validity are good (Hair et al. 2010). Following this line of reasoning, the 6-items were grouped as one variable called “Personality trait” taking the average of all items. This construct is used in CBC analysis as active covariate, where higher (lower) score indicates higher (lower) probability of an individual being associated as a maximizer (satisficer).

3.4.2 Conjoint analysis

The choice based conjoint analysis was performed using Latent Gold 4.5. Firstly, the aggregate model was analyzed to determine which attribute of a shopbot is on average more important. Thereafter, multiple-class analyses were performed to group respondents into segments

depending on their preferences and their demographical characteristics. The range of solutions was limited from 1 to 5 classes to reduce complexity as well as to have a realistic number of segments for managerial implications. Respondents’ part-worths were estimated, holding out the validation profile. The final estimated part-worths model, which will be discussed in the next section, is a result of an evolutionary, empirical estimation approach, were details such as

covariates and number of classes are subsequently included stepwise. This allows the researcher to reconcile simplicity with completeness and identify structural changes in the parth- worth utilities (Leeflang et al. 2000).

3.4.3 Ordinary Least Square

3_{A reliability test of 2 and 3 components was also performed, but it gave a Cronbach alpha below 0.6 which}

(27)

21

(28)

22

4.0 Results

4.1 Descriptive statistics

Demographic characteristics of the participants in the present study are shown in table 3.

Nationality distribution showed a bias toward European respondents when compared to the other nationalities. Likewise is the education distribution, where 29% and 55% of the respondents respectively have a Bachelor and Master degree. In terms of gender, there are 58% females and 42% males in the study and the age ranges from 18 to 60 years. The largest group is those between 25 to 29 years old, with 17.6% of total 202 respondents being 25 years old. With regards to occupation, most respondents are employed (38.61%), student (32.67%) and student with part-time job (26.73%).

Table 3: Demographic characteristics of the participants in the CBC analysis

% %

Age (years) <20 0% Education High school 9%

20-24 32% MBO 1%

25-29 67% HBO 3%

30-34 9% Bachelor 29%

>35 5% Master 55%

Gender Female 58% MBA/Phd 2%

Male 42% Nationality European 82%

Occupation student 33% Asian 9%

student with part-time job 27% North-American 4%

employed 39% South-American 1% unemploy 1% Africa 1% retired 1% Other 2% Demographic characteristics Demographic characteristics Sample size = 202

The personality trait score ranged from 1 to 7 with an average mean of 4.4. Figure 4 indicates on average that 61% of respondents scored above 4 in a scale from 1 to 7, which indicates that the largest proportion of respondents have relatively high maximizing tendencies.

Figure 4: Percentage people with the corresponding personality trait score

(29)

23

Table 4 below shows how respondents prefer each attribute by rating to what extent they agree or disagree with the statements. A high percentage of people, respectively 86% and 80% (above neutral) prefers smart search criteria and low waiting time. Overall, the results indicate that shopbots with smart search criteria and low waiting time are generally more preferred than number of alternatives, price chart displayed and bundle promotion, reflected in the higher mean score relative to the other attributes. This is found to be in line with the results obtained from CBC analysis, which will be discussed in a later section. An interesting notice is that 65% (above neutral) prefers a shopbot that displayed price chart, which was found to be opposite from the CBC analysis. The results may differ due to the different way of questioning and will be further elaborated in the discussion section.

Table 4: Attribute preference rating

I would prefer… "smart" over "simple" Low waiting time Many number of alt. Discount bundle Price chart

Freq. % Freq. % Freq. % Freq. % Freq. %

strongly disagree 5 2% 5 2% 9 4% 9 4%

disagree 5 2% 8 4% 16 8% 9 4% 21 10%

somewhat disagree 12 5% 7 3% 42 20% 20 9% 21 10%

neither agree or disagree 17 7% 23 11% 34 16% 48 21% 22 10%

somewhat agree 29 12% 33 16% 36 17% 55 24% 49 23%

agree 62 25% 53 25% 40 19% 51 22% 41 19%

strongly agree 122 49% 83 39% 39 18% 38 17% 50 23%

Mean 6.01 5.65 4.68 4.9 4.9

Total 247 100% 212 100% 212 100% 230 100% 213 100%

4.2 Choice-Based Conjoint Analysis 4.2.1 Aggregate results

Table 5: Utility parameters and relative importance of the aggregate model

Attribute s Le ve ls Wald Utilitie s Re lative importance (%)

Type* Simple 139.6518 -0.4407

Smart 0.4407

Waiting time* 20 sec 130.0636 0.4503

50 sec 0.1693

80 sec -0.6197

Number of alternatives* 10 37.4159 -0.2880

50 0.2156

100 0.0724

Price information* Combo 106.4136 0.4800

Price chart only -0.4476

Total price only -0.0324

Discount Bundling* 50% cross-bundling discount 80.7622 0.3976

No cross-bundling discount -0.3976

* p<0.001, Bold: highest score in utility level for each attribute correspond to higher preference

19% 22% 12%

(30)

24

Based on Wald statistics, table 5 indicates that all five attributes are highly significant (p-value<0.001). On aggregate level, waiting time was the most important attribute among the 202 respondents included in the choice based analysis. Its weight of importance relative to other attributes is 26%. The next important attribute is price information, followed up by type and bundling recommendation. These three attributes have a weighted importance of respectively 22%, 21% and 19%, which suggest they have a rather equal importance in general. The least important attribute was number of alternatives (12%). The bolded parameter estimates indicates the attribute levels with the highest utilities. With this respect, the shopbots that yields the highest utility in general are the once that provides smart search criteria, moderate number of alternatives, combination of total price and price line chart and bundle discount with the lowest amount of waiting time.

4.2.2 Model fit Aggregate Model

Figure 5: Aggregated part-worth utilities for ordered values

Both waiting time and number of alternatives were checked for potential linearity due to its ordered nature. Figure 5 shows a potential linearity effect in waiting time, whereas number of alternatives shows a clear part-worth utility. Therefore, only waiting time is checked to examine the improvement of model fit. Changing the attribute “waiting time” to a linear function did not improve the model fit (Table 6).

Table 6: Model fit aggregate model

Therefore, all five attributes are treated as nominal variables as a better fit between observed and predicted values was identified by the lower Bayesian Information Criterion (BIC), Consistent Akaike Information Criterion (CAIC), Akaike Information Criterion (AIC) and AIC3, which are log-likelihood (LL) measures (Vermunt and Magidon 2005) (Table 6). Furthermore, the hit rate which is the percentage of corrected classification of observed and predicted choice on aggregate level, is 57% (218+169+303/1212). Although the hit rate is moderately good, its predictive performance is better than a random prediction of choices where the predicted hit rate would be 33.33% (1/3 options). Therefore, relative to the random choice prediction, this model has a better predictive power.

Aggregate Model df BIC AIC AIC3 CAIC

(31)

25

4.2.3 Hypotheses testing on aggregate results

Hypotheses H1to H5 were all supported. In terms of type of shopbot, on average “smart” search criteria gives higher utility than “simple” search criteria, which supports H1. Waiting time does have a significant negative effect on the level of utility, indicated by a downward somewhat linear line. Hence, H2 is supported. The utility for a consumer is significantly highest for

moderate number of alternatives. The part-worth utilities for number of alternatives suggest that 50 alternatives have the greatest effect on utility, in support of H3. A combination of total price and price chart displayed gives as expected the highest utility. Hence, H4 is supported.

Furthermore, Cross-discount bundling yields a higher utility than no bundling in consumers’ choice preference of a shopbot, which supports H5. In sum, on an aggregate level, the best preferred option on average would be a smart shopbot with the lowest waiting time, which provides combined price information, cross-bundling discount and 50 alternatives.

4.2.4 Selection of number of segments and model fit

As in every market, different subgroups of segments may be present. This section seeks to identify the potential number of segments with respect to the preference for a shopbot by estimating a latent class model. The range of solutions is limited from1 to 5 classes due to the relative low number of respondents and to reduce complexity. Based on an exploratory approach, stepwise including/excluding covariates age, gender, occupation, education and nationality did not have significant effect on classifying the segments. Therefore, the model comparison as depicted in table 7 and table 8 is based on personality trait as the only active covariate.

Table 7: Fit Criterion of 1 to 5 class model

Proportion size pe r class

Model df BIC(LL) AIC(LL) AIC3(LL) CAIC(LL) Class. Err. 1 2 3 4 5

1-Class 194 2220,2924 2193,8263 2201,8263 2228,2924 0

2-Class 184 2199,245 2139,6957 2157,6957 2217,2446 0,1064 63% 37%

3-Class 174 2204,5236 2111,8921 2139,8921 2232,5236 0,1454 40% 35% 24%

4-Class 164 2216,5595 2090,8453 2128,8453 2254,5595 0,1431 50% 19% 17% 14%

5-Class 154 2235,4351 2076,638 2124,638 2283,4351 0,1564 27% 22% 21% 17% 14%

To determine how many segments are appropriate, an assessment of the model fit based on likelihood and classification will be discussed. As fit of the model always increases when additional parameters such as classes are added, the “relative fit”, which is the likelihood of observing the observed values given the model parameters corrected for number of model

(32)

26

is indicated by a lower value of the goodness-of-fit measurements. The widely used model comparison fit criteria BIC and AIC which penalizes for number of estimated parameters show contradicting model performance. Although a two-class model gave the lowest BIC, CAIC and classification error among classes, the AIC and AIC3 criteria suggest that the five-class model is the best model. The question arises as to which of the two model criteria are best in practice. For the purpose of providing mangers good future prediction which AIC and AIC3 emphasize on (Dziak et al. 2012), the 5 class model was chosen. Classification is best if subgroups are strongly separated (Vermunt & Magidson 2005). It determines how well the model can classify each respondent into classes given the observed data. Although the classification error is lower for the 2-class model than the 5-class model (table 6), an exploration into the class differences in terms of attribute importance favored the 5 class model. First of all, for the 2-class solution the order of the attributes’ relative importance is rather similar, which makes these two classes less

distinguishable and interesting in comparison to the 5-class model where there is clear distinction of attribute preferences. Second, for the two class model, the Wald (=) statistics for waiting time and price information is not significant (p=0.54; p=0.41; p=0.05 respectively), which suggests that the preference between classes in terms of the attributes waiting time and price information are rather equal. This is opposite for the 5-class model where all attributes preferences between classes varies significantly (p<0.05). Finally, only the 5-class model shows that personality trait has a significant effect in classifying the segments. Thus, the 5-class model provides a more interesting use for the purpose of this study.

Table 8: Comparing 2 class model with 5 class model

Wald p-value Wald (=) p-value Wald p-value Wald (=) p-value Attributes

Type 45,4996 1,30E-10 37,5225 9,00E-10 39,1037 2,30E-07 27,9088 1,30E-05 Waiting time 103,5900 1,70E-21 1,2486 0,54 44,5396 2,60E-06 27,9541 0,00048 Number of alternatives 36,4524 2,30E-07 17,5440 0,00016 37,0372 5,60E-05 22,0705 0,00480 Price information 102,1438 3,40E-21 1,8075 0,41 66,1985 2,40E-10 40,4210 2,70E-06 Bundling 67,9256 1,80E-15 7,3434 0,0068 62,7973 3,20E-12 34,9920 4,70E-07 Covariate

Personality Trait 0,1015 0,75 12,1346 0,016

Class 1 Class 2 Class 1 Class 2 Class 3 Class 4 Class 5

Class size 63% 37% 27% 22% 21% 17% 14% Type 9% 14% 11% 16% 1% 21% 40% Waiting time 29% 36% 17% 52% 13% 12% 10% Number of alternatives 8% 5% 13% 8% 17% 24% 24% Price information 25% 23% 53% 6% 18% 20% 9% Bundling 28% 21% 5% 19% 51% 23% 17% 2 Classe s 5 Classe s

(33)

27

Overall, the predictive power of classifying percentage of observed choices and predicted choices across 5 segments is 75% (274+285+348/1212) which is better than both the aggregate (57%) and the random choice (33.33%) models.

4.2.5 Segment results - 5 Class Solutions

Figure 6 illustrates the relative importance of each attribute and table 9 shows the utility of each attribute level. All five attributes are highly significant (Wald, p<0.001), also across the five classes (Wald (=), p<0.001). The findings suggest that the preferences are quite different from class to class. Figure 6 shows that price information, waiting time, bundling option, number of alternatives and type are the relative most important attributes for respectively class 1,2,3,4 and 5.

Figure 6: Relative important attributes for each class

0% 10% 20% 30% 40% 50% 60%

Class 1 Class 2 Class 3 Class 4 Class 5

Type Waiting time

(34)

28 Table 9: Utilities and relative importance of each segment

Class 1 Class 2 Class 3 Class 4 Class 5

N(%) 27% 22% 21% 17% 14% R² 31% 56% 25% 55% 56% Re lative importance (%) Type 11% 16% 1% 21% 40% Waiting time 17% 52% 13% 12% 10% Number of alternatives 13% 8% 17% 24% 24% Price information 53% 6% 18% 20% 9% Bundling 5% 19% 51% 23% 17%

Estimate d parame te r value s β z-value β z-value β z-value β z-value β z-value

Simple -0.5787 -2.4184 -1.3167 -2.9808 0.0343 0.127 -4.4495 -2.4888 -4.3082 -4.4131 Smart 0 . 0 . 0 . 0 . 0 . 20 sec 0.8927 3.2971 4.2118 4.2261 0.3498 1.2142 1.7364 1.6316 0.9066 2.0198 50 sec 0.3105 1.2774 3.0212 3.4078 0.4523 1.6629 2.5485 1.599 1.0959 2.4471 80 sec 0 . 0 . 0 . 0 . 0 . 10 alternatives -0.6719 -2.8457 -0.3891 -0.8901 0.5386 1.6202 -5.0572 -1.9575 -1.5473 -2.6480 50 alternatives -0.1462 -0.5667 0.2251 0.6798 0.5705 2.1135 -1.0742 -1.3883 0.9857 1.8120 100 alternatives 0 . 0 . 0 . 0 . 0 . Combo 0.6071 3.2205 0.3425 0.8739 0.6142 2.1215 3.5778 1.9517 -0.9737 -1.4677

Price chart only -2.1024 -4.8457 -0.1109 -0.2531 0.0725 0.1787 -0.5601 -1.1244 -0.5631 -1.0366

Total price only 0 . 0 . 0 . 0 . 0 .

50% cross-bundling discount 0.2633 0.6877 1.5016 3.0948 1.7647 6.2585 4.8339 1.9263 -1.8536 -2.5283

No cross-bundling discount 0 . 0 . 0 . 0 . 0 .

Intercept -1.8317 -1.4914 -3.7955 -2.6138 -2.4894 -1.8454 -4.7055 -3.1442 0 .

Covariate: Personality trait 0.6177 1.9728 1.0147 2.9394 0.7116 2.1341 1.1537 3.2684 0 . a. The parameters are set to zero because it is redundent (dummy last code)

b. Wald and Wald(=) associated p-values for all parameters are all highly significant (p<0.001), c. Intercept and personality trait are both significant on 5% significant level.

d. Bold=│z-value│>1.97 = p<0.05

(35)

29

and gender for certain classes. These will be further discussed in the section of profiling and interpretation of each segment.

4.2.6 Predictive validity

To validate the predictive power of both the aggregate and 5-class model, the individual

parameter coefficients estimated were utilized for calculating the predictive validity of the choice based conjoint analysis using the hold out sample. The predictive power identified as the hit rate is defined by dividing the sum of the correct estimated choices by the total observed choices. Table 10 compares the in-sample and hold-out sample hit rate for both the aggregate and 5-class models. The 5-class model gave the best prediction hit-rate of 75% and 61% for both the

estimated and holdout samples respectively (appendix 6). A hit rate of 61% means that in 61% of the cases, the predicted choice was equal to the actual observed choices for the 202 respondents. Both models predicts better than a random prediction of 33.33% (1/3 options) (table 10). Thus, in comparison to the random choice prediction, the 5-class model has a better predictive power4.

Table 10: Hit-rate across 2 models

Model In-sample Out-sample

Aggregate 57% 34%

5 class model 75% 61%

4

(36)

30

4.2.7 Profiling and interpretation of the segments Class 1 – Information oriented users

Figure 7: Relative Importance for "Information Oriented users"

The beta effect estimates for class 1 suggest that segment 1s’ preferences are influenced in a positive way by shopbots for which type = smart, waiting time =20 sec, number of alternatives = 100 alternatives, price information = combination of price chart and total price displayed,

bundling = with 50% cross-bundling discount. Clearly, price information has the strongest effect on class 1’s preferences reflected in the highest relative importance of this attribute (53%). The other attributes have more or less equal effects except bundling option, which possess the lowest importance. The probability that class 1 differs from class 5 in terms of personality trait is significant (p<0.05). The significant positive utility score of personality trait (βPT=0.6177, p<0.05) suggests that class 1 has a higher likelihood to be a maximizer than class 5. However, because it does not possess the highest nor the lowest parameter in terms of personality trait, it may be recognized at somewhere between a maximizer and satisficer, as the highest percentage of the members (26%) scored moderately in the personality trait scale (1-7). Class 1 clearly indicates its value for combination of both price chart and total price displayed. Furthermore, as there seems to be a positive relation with number of alternatives provided, this class may be profiled as “information oriented users”. Class 1 is the largest segment (27%) consisting of more females than males, 70% and 30% respectively.

(37)

31

Class 2 – Efficient user who value cross-discount bundling

Figure 8: Relative Importance for "Efficient users"

(38)

32

Class 3 – Discount bundling oriented users

Figure 9: Relative Importance for "Discount bundling oriented users"

Class 3 is the third largest segment (21%) consisting mainly of students, slightly more male than female (52% vs. 48%).Their preference of a shopbot is positively influenced by which type= simple, waiting time =50 sec, number of alternatives = 50 alternatives, price information = combination of price chart and total price displayed, bundling = with 50% cross-bundling discount. For this group, providing cross-bundling 50% discount in relation to no bundling clearly has a positive effect on utility, which is also reflected in the weighted importance of the attribute (52%). A reduction of 100 to 50 numbers of alternatives gives a significant higher shift in mean utility in comparison to a reduction of 100 to 10 numbers of alternatives. Hence, there is no real difference in utility between 10 and 50 numbers of alternatives, but 100 numbers of alternatives is clearly not preferred. With respect to price information display, there is a

significant positive effect on utility between displaying only total price and combination of total price and price line chart (βprice_combo=0.6142, p<0.05). Although price chart only has a positive significant effect on utility, its effect is not significantly different from displaying only total price. Interesting about this segment is that it is the only one that value simple search criteria. Class 3 is more likely to be classified as a maximizer than class 5 and slightly more likely than class 1, but less likely for class 2 and 4. In general, due to its high preference of discount bundling, this group is profiled as “discount bundling oriented users” with a touch of maximization personality trait.

(39)

33

Class 4 – Optimizing users

Figure 10: Relative Importance for "Optimizers"

Class 4s’ preference of a shopbot is positively influenced by which type= smart, waiting time =50 sec, number of alternatives = 100 alternatives, price information = combination of price chart and total price displayed, bundling = with 50% cross-bundling discount. All attributes except for waiting time seem to have equal effect on utility for class 4. However, number of alternatives provided is most valued given the highest relative importance of 24%. The relative linearity of this attribute suggests that for this group, there is an increasing positive effect with number of alternatives provided. The next important influential effect on preference is cross-discount bundling offer followed by smart type and price combination information. It seems that class 4 is willing to trade-off waiting time with more information and solutions because highest utility attached to waiting time was 50 seconds. They find smart shopbots that provide 100 alternatives, combined price information and 50% bundling discount more preferable. Interestingly, this group has the highest probability of being optimizers and are significantly different from class 5 (β=1.1537, p<0.05). The relative equal importance of attributes also confirms their maximizing tendencies. Thus, this group can best be profiled as “the optimizers” who typically are characterized with more females than males, 63% versus 37% respectively.

User’s preferences in comparison shopping websites