The most optimal amount of choice sets in a conjoint analysis and the analysis of the drivers of willingness to respond

(1)

The most optimal amount of choice sets in a conjoint analysis

and the analysis of the drivers of willingness to respond

What is the influence of forcing respondents on their choice behavior?

By

(2)

1

The most optimal amount of choice sets in a conjoint analysis

and the analysis of the drivers of willingness to respond

What is the influence of forcing respondents on their choice behavior?

By

Iris Kluin

Faculty of Economics and Business, University of Groningen MSc Marketing Intelligence and Marketing Management

MSc Thesis January 2016 Pioenstraat 107 9713 XT, Groningen (06) 23690575 i.kluin@student.rug.nl S1771531

1st_{Supervisor: dr. F. (Felix) Eggers}

(3)

2

Abstract

Choice- based experiments are nowadays very popular to reveal what characteristics consumers prefer the most on a product or service. In order to get the most optimal results from respondents, the survey should contain the right amount of choice sets. According to previous literature, too many choice sets can create boredom and fatigue for respondents which can result in a change in the attribute importance. Other articles conclude too few choice sets will lead to incomplete information. A voluntary situation can solve boredom and fatigue of taking a survey, this creates more willingness to participate. A voluntary design allows respondents to stop the survey on the moment they do not want to answer any additional questions anymore. According to Regner (2014), there are four motivational drivers: fairness, reciprocal concerns, self- image concerns and norm conformity. These drivers increase the willingness of people to participate in voluntary situations according to the literature.

In this paper, the effect of the number of choice sets on the attribute importance is tested in a choice experiment for flight trips in Europe with a forced and voluntary design. Data is collected by observing 195 respondents. First, the influence of the four motivational drivers of a voluntary setting is tested on the number of choice sets answered by respondents. This effect is measured through an ordinary least squares regression. Second, the influence of the number of choice sets on the attribute importance is measured in a forced design as well as a voluntary design. Therefore, a conjoint analysis, generated in the program Latent Gold, is used.

The findings of this study regarding the number of choice sets answered in a voluntary design supports the literature since this results in less choice sets answered by respondents. Moreover, respondents with a higher level for norm conformity will answer more choice sets. Additionally, the number of choice sets has an influence on the respondents’ choice behavior, since this changes during the survey. Finally, the voluntary design, in comparison with the forced design, results in a decrease in differences in the attribute importance. Respondents diminish changes in their choices towards airline, amount of stops, baggage and ticket price.

(4)

3

Preface

This thesis is written to accomplish my double- profile Msc Marketing Intelligence and Marketing Management at the University of Groningen. In this report I investigated at first, the influence of a voluntary design on the number of choice sets answered. Second, the influence of the number of choice sets in a voluntary and forced design on the attribute importance is researched. I would like to thank my first supervisor dr. F. (Felix) Eggers for his helpful guidance throughout the semester and my second supervisor prof. dr. J.E. (Jaap) Wieringa for his supervision. Additionally, I want to thank dr. F. (Felix) Eggers also for the use of his software Preference Lab.

(5)

4

1. Introduction

Preference measurement is important since these analyses accomplish sustainable customer information, a valid ground for predicting their choice behavior and buying decisions (Eggers and Sattler, 2011). Within the area of marketing, the conjoint analysis is one of the most popular methods to measure preferences (Bradlow, 2005). In a choice-based conjoint (CBC), respondents choose their most desirable product composition from a few alternatives (Eggers and Sattler, 2011). The CBC has become a very common method to measure the preferences of respondents efficiently (Eggers and Sattler, 2009). Within a CBC, consumers have to choose between different options, however, the researcher is required to make other choices. When a researcher decides to do a CBC study, he has to determine how many choice sets he wants to ask the respondent (Johnson and Orme, 1996). Asking too few choice sets will result in a reduction of precision (Johnson and Orme, 1996). However, too many choice sets may result in biased or unclear results. It can become very monotonous if respondents have to answer choice sets repeatedly, hence there is the possibility of fatigue or boredom after answering multiple questions. Respondents’ fatigue and low attention span can change their behavior and the choices they make during the survey, which is undesirable (Eggers and Sattler, 2011).

Bech, et al. (2011) investigated these effects and found that indeed respondents’ choice behavior changes while answering multiple choice sets. Multiple articles about the amount of choice sets a researcher can ask a respondent are published, however not all reach the same conclusion. For instance, Carlsson and Martinsson (2007) mention that there is no change in the choice behavior of respondents while answering more choice sets. Furthermore, Bech, et al. (2011) believe that there is no consensus on the right amount of choice sets within a survey. Chung, et al. (2011) conclude that surveys with different amounts of choice sets results in differences in respondents’ choices, since a change in attribute importance is observed. Due to conflicting findings, more research is needed into this subject. As a researcher, it is useful to know the exact number of choice sets one can ask respondents. This will lead to the most optimal and reliable results, which points to the first research question: ‘what is the most optimal number of choice sets you can ask your respondents?’.

(8)

7 consumers have maximum control over the price they pay (Kim, et al., 2009). People are willing to take action voluntarily, therefore the research explores if respondents are willing to voluntary answer and keep their focus while answering all choice sets. This voluntary approach will be contrasted with a mandatory design, where respondents are forced to continue and finish answering all the choice sets. This leads to the second research question: ‘does the change in decision behavior

still occur when using a voluntary design?‘.

This research illustrates that there are mixed findings in relation to the optimal amount of choice sets and the influence of the number of choice sets on the attribute importance. Since there is not much theory on this subject, more research is needed into how the number of choice tasks a researcher asks respondents can influence choice behavior. Additionally this enhances the knowledge and outcomes of the conjoint analysis. Therefore, this research is more focused on empirical evidence and less on theory. Second, a forced and voluntary situation will be compared in terms of their effect on the attribute importance and the choice behavior of respondents. This helps to optimize and improve the conjoint analysis method.

This paper is structured as follows. First, the theoretical framework is provided, which addresses a literature review with prior research; this leads to the research questions. Second, based on the literature review, a conceptual model can be built. This will graphically display the relationships found in the literature, and is the guideline with which empirical data is gathered. Hypotheses are formulated after the conceptual model is presented. Third, the research methodology is explained in chapter four. This contains the choice of research and the data collection method. In addition, the results of the analysis will be presented in chapter five. These outcomes will be discussed in comparison with the existing literature in the following chapter. Finally the conclusion with implications and limitations will close this study.

2. Literature review

In this chapter, relevant literature and prior research for this study are presented. This consists of two parts. First, literature in relation to the effect of the number of choice sets on respondents’ behavior. Second, the voluntary design is explained with the Pay- What- You- Want pricing method. Based on this literature a conceptual model is drawn in chapter 3.

2.1 Number of choice sets

(9)

8 Respondents can get tired or bored, and the chance of not paying attention when a survey is too long increases. This is an important issue because their answers may become of little value when respondents get bored and this may impact the results (Johnson and Orme, 1996). Researchers want to collect as much data from each respondent as possible, however the survey does not have to be too long. Resulting in the question: what is the most optimal amount of choice sets? Johnson and Orme (1996) found some contrary results to what one might expect. Their findings stated that the answers given later in the survey are more reliable than the answer given earlier in the survey. Johnson and Orme (1996) concluded that the reliability moves upwards as the interview progresses, until ten tasks. After the tenth task there is a modest increase in reliability till the 20th_{task. Additional}

findings from this article provides that the answers of respondents will change in a way of a change in attribute importance. The attribute brand becomes less important, and price become more important when the survey advanced according to Johnson and Orme (1996).

(10)

9 quite small, however they are significant. According to their results 16 treatments are sufficient, although with 24 and even 32 treatments the outcomes will always be improved. They stated that once 16 treatments are used, work may well pay off in the larger sample sizes, especially in ensuring that there are enough choices of each alternative in the choice set (Hensher, et al., 2001). Moreover in the article of Hess, et al. (2012) they came up with clear findings about fatigue effects in a larger number of stated choice surveys. The engagement with the survey changes across different amount of choice tasks (8, 15 and 16) is stated in their results. The difference with other studies is that they used multiple datasets, which makes it more easy to generalize their results (Hess, et al., 2012). Lastly, Swait and Adamowicz (2001) made a model that at the same time considers the complexity of tasks, the effort consumers practice, the choice ability and their choice. Task complexity has an impact on the choice behavior of consumers and that choice behavior changes after 8 choice sets (Swait and Adamowicz, 2001).

All these articles show that the amount of choice sets do have an effect on the choice behavior of respondents. In contrast towards these articles the article from Louviere, et al. (2013) stated the opposite. The impact of the number of choice sets on the choice behavior is little and not always happening (Louviere, et al., 2013). In addition to this article, Carlsson and Martinsson (2007) also found other results. They did not find an effect between the amount of choice tasks and the results of the behavior of the respondents. Two situations were tested, one of 12 choice sets and one of 24 choice sets. The utility did not change during both the surveys in both situations (Carlsson and Martinsson, 2007). Concluding out of these findings it can be stated that there are mixed findings about this subject and more research is needed.

(11)

10 Table 1

Prior research into the number of choice sets in conjoint analysis

Authors Research context Number of

choice sets

Utility Evolution

Bech, Kjaer, and Lauridsen 2011 Dental services 5, 9, and 17 Yes

Carlsson and Martinsson 2007 Unplanned power outages

12 and 24 No

Chung, Boyer, and Han 2011 Beef 1- 20 Yes

Hensher 2006 Commuter trips 6, 9, 12, and 15 Yes

Hensher, Stopher, and Louviere 2001

Flight trips 4, 8, 16, 24, and 32 Small

Hess, Hensher, and Daly 2012 Various transportation studies

8, 15, and 16

(not experimentally controlled)

Yes

Swait and Adamowicz 2001 Orange juice 16 Yes

Johnson and Orme 1996 Various, across 21 data sets

8 -20 Yes

2.2 Voluntary design

(12)

11 for it. Some consumers act like this to be fair towards companies (Jang and Chu 2012). They concluded that there are customers who behave in a fair way towards a company, even if they are not committed to do so (Jang and Chu, 2012).

Transferring these findings to a survey setting, it can be assumed that respondents will behave in a fair way, with the result that they voluntary help a company or researchers’ experiment. In order to help the researcher, respondents will answer fairly and honestly in a voluntary design. Assuming that people spend time to voluntarily take action for researchers. The same principle as paying voluntarily in restaurants because, time is money.

3. Conceptual model

In this chapter a conceptual model is drawn, based on prior research. This conceptual model is a representation of the research that will be conducted. Hypotheses from this literature and the conceptual model are formulated in the next part of this chapter.

Figure 1

Conceptual model shows the influence of the number of choice sets on attributeimportance

(13)

12 on the number of answered choice sets. Second, the effect of the number of choice sets on the attribute importance will be investigated. Previous research conclude that answering more choice sets will have an effect on the choice behavior of respondents. Johnson and Orme (1996) concluded that the attribute brand becomes less important, and price becomes more important when the amount of choice sets increases. Finally, this effect will be moderated by a voluntary design. In this situation respondents are not forced to finish the whole survey, however they voluntary decide whether to continue till the end or stop earlier. This moderating effect will test if the number of choice sets respondents are willing to answer in a voluntary design still cause the same changes in the attribute importance.

3.1 The effect of the voluntary design on the number of choice sets answered

The voluntary design consists of a survey in which the respondent can quit during the survey. Thereby the respondent is not forced to finish it and is able to skip earlier to the end. Equal as in a PWYW settings, respondents voluntarily take action out of a sense of fairness (Jang and Chu, 2012). Most of the guests pay always and above the minimum, therefore they are willing to spend money on it voluntarily. However, it is mostly lower than the average price (Riener and Traxler, 2012). In the PWYW setting, consumers are willing to pay voluntarily for the products or services to stay fair towards themselves and the company (Jang and Chu, 2012). These consumers are free to decide what they want to do, and they choose to participate and answer the choice sets. Regner (2014) stated that consumers pay significantly above the minimum price. Therefore, it can be concluded that participants in a voluntary design are willing to answer choice sets, however they will answer less choice sets than the respondents in an obligatory situation. Since people often pay less than average in a PWYW setting. This leads to the first hypothesis:

Hypothesis 1 Voluntary design leads to a significant lower number of choice sets answered than in

the forced design.

(14)

13 behavioral success factors for PWYW pricing. Thus, people who have a high level of fairness will help more and offer more of their time to participate. Therefore, the hypothesis is written as;

Hypothesis 1a A stronger level of fairness will lead to more choice sets answered than people with a

lower level of fairness.

The second motivational driver of the PWYW policy, is reciprocal concerns. In a model of sequential reciprocity, customers with concerns for reciprocity will appreciate this kind of voluntary payments and tend to pay more than necessary, while customers with no of low concerns for reciprocity will not (Dufwenberg and Kirchsteiger, 2004). Therefore, people who have a higher concern for reciprocity, are able to participate and answer choice sets in a voluntary setting easily. Reciprocity is described by Machado and Sinha (2013, p.4) as, “the actions of individuals that are often governed by how they are viewed by others as well as their concerns about how their choices affect the behavior of other towards them”. Some people reciprocate to kind intentions that are expressed by kind actions (Schmidt, et al., 2015). Therefore, consumers answer as kind of kindness in a voluntary setting. The participant will answer more choice sets in a voluntary setting when their reciprocal concerns are higher. This will lead to the second hypothesis of the motivational drivers;

Hypothesis 1b A higher level of reciprocal concerns will lead to more choice sets answered than

people with a low level of reciprocal concerns.

Gneezy, et al. (2012) conclude that self- image concerns are an important determinant of voluntary payments (Regner, 2014). Therefore, consumers who choose to pay voluntary make these decisions since they feel that paying is the right thing to do. When people participate in a voluntary situation, they choose to do this to maintain an individuals’ positive self- image (Gneezy, et al., 2012). Jang and Chu (2012), found that self- signaling is one of the two principal underlying mechanisms of PWYW. Self- signaling to signal to themselves that they act fair and honest (Jang and Chu, 2012). Therefore, consumer do pay more when they care more about their self- image (Jang and Chu, 2012). Regarding participation in a voluntary survey, respondents with a high concern about themselves are more willing to answer choice sets than people with a low concern of self- image. Therefore, in a voluntary setting people that want to keep a positive self- image are willing to participate and help the researcher by answering choice sets. This results in the third hypothesis;

Hypothesis 1c A higher level of self- image concerns will lead to more choice sets answered than

people with a lower level of self- image concerns.

(15)

14 the researcher, since people feel bad when violating the norm. Norm conformity is next to self-signaling, the other principal underlying mechanism of PWYW (Jang and Chu, 2012). Therefore, respondents who value norms high are more likely to answer more choice sets than respondents with lower level of norm conformity. Varman and Costa (2008) wrote that social norms may determine consumer behavior (Jang and Chu, 2012). They concluded that a fair descriptive norm will lead to more fair behavior of consumers. When respondents are asked to participate under strict norms, they will be more willing to answer more choice sets. These findings lead to the fourth and last sub hypothesis:

Hypothesis 1d A higher level of norm conformity leads to more choice sets answered in a voluntary

setting.

3.2 The effect of number of choice sets on attribute importance

Multiple articles stated that an increase in complexity and length of a survey will lead to a higher error variance in choice experiments (Chung, et al., 2011). A higher amount of choice sets will lead to a small change in decision behavior for flight choices among the respondents according to Hensher, et al. (2001). When the amount of choice sets is larger, this results in more observations and more information, however there is also more room for inaccuracy due to boredom and fatigue (Bech, et al., 2011). Contrary to this, when there is a lower amount of choice sets used, this can lead to incompleteness and missing information (Bech, et al., 2011; Johnson and Orme, 1996). Respondents get tired or bored and answering repeatedly questions will require a lot of a respondents’ effort (Bech, et al., 2011). There is a chance that their decision making behavior changes, which can result in a change in the attribute importance. Bech, et al. (2011) concluded that there is no one best appropriate number of choice sets. However, mostly the amount of choice tasks is around eight till sixteen (Bech, et al., 2011). This concludes that above sixteen the tiredness and boredom of respondents increases and does have an impact on the attribute importance. Hence in the article is concluded that an increase in the number of choice sets will lead to a higher price sensitivity (Bech, et al., 2011). This confirms the statement of Johnson and Orme (1996), who concluded that price becomes more important when the size of choice sets increases. Carlsson and Martinsson (2007) did research with two conditions, one with 12 choice sets and one with 24 choice sets, however did not find a significant difference in the respondents’ behavior. Due to the mixed findings, more knowledge is needed. The expectations are that the attribute importance changes when using larger choice set sizes than smaller ones. All these references together will lead to the second hypothesis:

Hypothesis 2 A higher number of choice sets leads to differences in attribute importance (price

(16)

15 3.3 The moderating effect of the voluntary design on the attribute importance

The voluntary design will create a feeling of freedom among respondents. Freedom to participate and answer carefully and unforced. This is the same as the PWYW pricing, because customers are prepared to pay voluntarily (Regner, 2014). Assuming in this research that people spend time to voluntarily take action for researchers. When respondents are forced to finish a survey I expect respondents to answer latter questions less carefully than the earlier questions. This is multiple times measured in prior research. Causssade, et al. (2004) concluded that there is an optimal amount of choice sets (9 or 10) and that more choice sets will lead to higher error variance. The more choice sets are asked or the more complex the survey is, the higher the inaccuracies in the results, and the more respondents change their behavior when making decisions (Johnson and Orme, 1996). This will lead to the utility evolution, which is not desirable in any research. Therefore, voluntary design might lead to a decrease in this change in decision behavior. In order to help the researcher, respondents will answer fairly and honestly in a voluntary design (Jang and Chu, 2012). Next to this there are probably more respondents who give honest and well overthought answers in a voluntarily situation than in a forced situation because they do not feel rushed or pushed. In the case of answering choice sets voluntarily, this will lead to better outcomes than when respondent are forced to answer. Finally this will lead to the third hypothesis:

Hypothesis 3 The moderating effect of using a voluntary design decreases the difference in attribute

importance.

4. Research methodology

In this chapter the method of data collection is described. First, the used method is explained with the experimental design, measures and procedure. Second, the steps of the analysis phase and the model selection is described carefully.

4.1 Method: Conjoint experiment

(17)

16 There are different types of conjoint analysis, the traditional conjoint, adaptive conjoint and the choice- based conjoint (Eggers and Sattler, 2011). The conjoint analysis is a decompositional approach, since respondents evaluates entire products by considering the product attributes and levels together (Eggers and Sattler, 2011). In this study, the choice- based conjoint is used, where the respondent is asked repeatedly to choose between different flight options and select their most preferred alternative. This approach is effective, since people make choices every day, this is a daily process, therefore people are used to do this (Eggers and Sattler, 2011).

4.2 Conjoint design

To study the influence of the amount of choice sets on the attribute importance, data is collected conducting an online choice experiment. This choice experiment is designed by the program Preference Lab. Data is collected by a survey with two different designs which are randomly divided among the respondents. These two different conditions are explained in paragraph 4.3; experimental design.

The research design is used from the article of Eggers and Sattler (2011), whereby the same attributes are used as in their design, however, some levels do change. Additionally, the number of attributes and levels are similar as in their design. The amount of attributes is five, with each three levels. Every attribute has three levels to avoid estimation error, since the reliability of each estimate drops, when more levels are integrated (Eggers and Sattler, 2011). Moreover, the levels are balanced across all attributes, to prevent the number of levels effect (Eggers and Sattler, 2011). The number of levels effect needs to be prevented since attributes with more levels result in relative higher attribute importance than attributes with lower levels, which can lead to bias in the results (Eggers and Sattler, 2011). The attribute level pair have to appear an equal amount of times, therefore the choice design needs to be balanced. Another criteria Eggers and Sattler (2011) handle is minimal overlap. Therefore, the presented alternatives are maximally different from one another in this experiment. For instance, by avoiding equal levels in one attribute. Therefore, all attributes and levels are presented one time per choice set. In table 2, the five attributes are presented with their descriptions and the three associated levels.

Table 2

Overview of attributes in the choice experiment

Attribute Description Levels

Airline Name of the airline Easy Jet KLM Air Berlin

(18)

17

Attribute Description Levels

Amount of stops Whether the flight is direct or with intermediary stops

Non

(direct flight)

One stop Two stops

Allowed baggage Weight of baggage 10 kg 15 kg 20 kg

Ticket price Price of the flight €49,- €99,- €149,-

As presented in the table, five attributes with each three levels are used in this CBC experiment, which means that there are 3⁵ = 243 possible flight options in total. This full factorial design is not an option to collect data, since no participant wants to evaluate 243 alternatives. Therefore, a fractional factorial design is used in this research, where a subset of all possible combinations of attribute levels is used. The program Preference Lab is used to create the survey and to collect data via a fractional factorial design.

4.3 Experimental design

Two conditions are used to set up this experiment; a forced and a voluntary design. Both conditions contain choice sets with five attributes, each with three levels. The difference between the conditions is the amount of choice sets respondents needs to answer and whether they can stop earlier during the survey. In the forced design, respondents are required to complete 24 choice sets. Hence, they cannot stop earlier and are obliged to continue till the end. By contrast, in the voluntary design, respondents are forced to answer the first 12 choice sets and are free to decide whether to stop or continue after this 12th_{choice set. These participants are able to stop when they do not want}

to answer additional choice sets anymore. Therefore, they can skip to the end of the survey, during choice set 13 till 24. Figure 2 shows a screenshot of the this page. This is conducted in order to check whether respondents answer less questions if they are allowed to stop and what the effects are of forcing respondents on their choice behavior.

Figure 2

(19)

18 In both conditions, respondents are able to answer 24 choice sets. Previous mentioned articles show that respondents fatigue and decreasing attention when a survey is too long, which changes the choice behavior (Eggers and Sattler, 2011). To avoid changing choice behavior, the survey design should include about 12-15 choice sets (Eggers and Sattler, 2011). According to Bech, et al. (2011) this amount is 8-16. This CBC experiment is developed to observe the behavior of respondents during a voluntary survey and the effect of enforcement. Results will show if respondents will change their behavior, whether there are differences between both conditions and what the effect will be of forcing respondents.

4.4 Measures

The four motivational drivers for answering in a voluntary setting are measured by different items. Table 3 shows these items and their sources. First, the construct fairness is measured by four items measured by Jesse, et al. (2009). Reciprocal concerns is measured by four items, three of them by McLure and Samer (2005) and one by Regner (2014). Third, the driver self- image concerns is measured with three items, these are based on the references’ experiments and conclusions of

Table 3

Items of the drivers that motivates respondents to answer voluntarily Drivers Items

Fairness F1 - If a friend wanted to cut in with me on a long line, I would feel uncomfortable because it wouldn’t be fair to those behind me (Jesse, et al., 2009).

F2 - In the fight against terrorism, some people’s rights will have to be violated (Jesse, et al., 2009). F3 - Justice, fairness and equality are the most important requirements for a society(Jesse, et al., 2009).

F4 - When the government makes laws, the number one principle should be ensuring that everyone is treated fairly (Jesse, et al., 2009).

Reciprocity R1 - When I receive help, I feel it is only right to give back and help others (McLure and Samer 2005). R2 - I know that other members will help me, so it's only fair to help other members (McLure and Samer 2005).

R3 - I trust that someone would help me if I were in a difficult situation (McLure and Samer 2005). R4 - If someone does something that is beneficial to you, I would be prepared to return a favor, even when this was not agreed upon in advance (Regner, 2014).

Self- image S1 - When I voluntarily pay for something, I do this to feel better (Regner, 2014). S2 - I donate money for charity to feel good about myself (Regner, 2014).

S3 - I donate more money for charity in public than in a private setting (Regner, 2014). Norm

Conformity

N1 - I tend to consider strongly what others believe is appropriate when I make a decision (Regner, 2014).

(20)

19 Regner (2014). Finally, the construct of norm conformity is measured by two items of Regner (2014).

4.5 Procedure

The survey is made in Preference Lab, within this program the choice sets are presented to the respondents. In appendix A, the whole survey is presented. First, introductory questions, to collect demographic information about the participants, are asked. These questions gather information about gender, age, household size, household income, educational background and employment status. To collect as much data as possible and avoid missing data, all questions of the survey needs to be fulfilled to continue the survey.

The second part of the survey measures the level of motivational drivers of the PWYW behavior of participants. Therefore, this part consists of different statements which measure; fairness, reciprocal concerns, self- image concerns and norm conformity. Respondents are asked to give their opinion about these statements with a Likert Scale of 0 (totally agree) to 4 (totally disagree). A 5-point Likert scale is used here, one of the most frequently used variation of the summated rating scales (Cooper and Schindler, 2008).

The conjoint experiment is the third part of the survey. Participants are told that they are free upcoming week and are able to book a flight in Europe. Additionally, the different alternatives are shown and participants needs to choose their most preferred flight option. Therefore, they are asked to select the option with their most preferred airline, destination, weight of baggage allowed, the amount of intermediary stops and finally the ticket price. Besides the question about their most preferred alternative, the survey added the question if this situation was real, whether respondents would book the flight (yes - no). A screenshot of a choice set is shown below in figure 3.

The two different designs are set up differently. Within the forced design, 24 choice sets are asked and all 24 should be answered. The voluntary design provides also 24 choice sets, of which 12 are mandatory and 12 are voluntary to answer. After the 12th_{choice set in the voluntary design, the}

survey gives the information that the participant is half way, and has to keep on going. However, the participant is allowed to skip additional choice sets, as depicted in figure 2. In the forced design after the 12th_{choice set, the survey provided the participant with information that they are halfway and}

they need to keep on going without the link to skip to the end of the survey. This information is provided as equal as possible for both conditions to treat every respondents in a similar way.

(21)

20 thereby the participants are mainly invited via Facebook and email messages. An English version and a Dutch version were presented.

Figure 3

Example of choice set question (Preference Lab 2015)

4.6 Model selection

In these paragraphs the model selection is explained for the different parts of the analysis. The ordinary least squares (OLS) regression model is described and sub sequential the model selection for the conjoint experiment is presented.

4.6.1 Ordinary Least Squares regression

(22)

21 coded. Concluding, an OLS regression is performed. To estimate parameters, the OLS regression is a common used method (Noreen, 1988). The dependent variable, number of choice set answered, is explained based on the different independent variables subsequently; fairness, reciprocal concerns, self- image concerns and norm conformity, in a forced and a voluntary situation. Formula 1, presents this model, which consists of the dependent variable, an intercept, which is the constant

𝛽

+ 𝛽

_{𝑛𝑜𝑟𝑚−𝑐𝑜𝑛𝑓𝑜𝑟𝑚𝑖𝑡𝑦}

𝑋

_{𝑛𝑜𝑟𝑚−𝑐𝑜𝑛𝑓𝑜𝑟𝑚𝑖𝑡𝑦}

+ ℇ

Where,

y = number of choice sets answered

𝛽

₀= the constant, the intercept ℇ = the error term

In order to perform a linear regression, items of the different drivers needs to be checked on collinearity and correlation (Belsley, et al. 2005). Additionally, to reduce the amount of variables, hence, to check if these items can be taken together, a factor analysis is performed (Hair, et al., 2009). Since the different items, presented in table 3, measure the four drivers of motivation, there is researched if the items can be taken together. This will result in four independent variables. This factor analysis is explained in the results; chapter 5.2.

4.6.2 Conjoint experiment

The second part of the model selection is about the conjoint experiment. Choices of consumers and participants are based on overall utilities of alternatives. Products are perceived as attribute bundles, and the utility of a product equals the sum of the utilities of the attribute levels. The utility of consumers (n) for product (i) is explained by formula 2 (Hair, et al., 2009).

Formula 2:

𝑈

_𝑛𝑖

= 𝑉

_𝑛𝑖

+ 𝜀

_𝑛𝑖 Where,

V = systematic utility component, rational utility ℇ = stochastic utility component, error term

(23)

22 shown in formula 3. Conjoint analysis systematically varies the attribute levels across alternatives and show different options to respondents, and by variation, this analysis tries to identify the systematic utility component (Hair, et al., 2009).

Formula 3: Systematic utility

𝑉

_𝑖

= ∑

𝐾_𝑘=1

𝛽

_𝑛𝑘

𝑥

_𝑖𝑘 Where,

V = systematic utility component, rational utility 𝑘 = number of attributes

x = dummy indicating the specific attribute level of product i β = part- worth utility of consumer n for attribute 𝑘

𝛽

𝑎𝑖𝑟 = part- worth utility for the airline, 𝛽𝑑𝑒𝑠𝑡 = utility for destination,

𝛽

𝑠𝑡𝑜𝑝 = utility for amount of

stops,

𝛽

_𝑏𝑎𝑔= utility for amount of allowed baggage and

𝛽

_𝑝𝑟𝑖

=

utility for price.

Concluding, the utilities can be measured for all alternatives. Based on these estimates the probability can be measured. The probability of choosing a certain flight (i) from choice set (J) is shown below in formula 6, a Multinomial Logit Model (Greene and Hensher, 2003).

Formula 6: 𝑝𝑟𝑜𝑏(𝑖|𝐽) = 𝐸𝑋𝑃 𝑉𝑖

∑𝑚𝑗=1𝐸𝑋𝑃 𝑉𝑗

𝑝𝑟𝑜𝑏(𝑖|𝐽) = probability of choosing a certain flight option

i = flight option J = choice set

(24)

23 importance will be affected by the range of levels which is chosen for each attribute. The formula 7 is shown below.

Formula 7: 𝑎𝑏𝑠( 𝑚𝑎𝑥 [𝛽𝑘]– 𝑚𝑖𝑛 [𝛽𝑘])

In the analyses, the effect of the number of answered choice sets on the attribute importance is measured. Additionally, the attribute importance is calculated for the forced and voluntary design, therefore these two conditions are compared with each other, hence, the effect of forcing respondents is investigated.

5. Results

First, this chapter presents a description of the sample and it shows the results of the factor analysis and regression analysis. These results show the effect of the four drivers of motivation to answer voluntarily on the number of choice sets answered. Second, this chapter contains the outcomes of the conjoint analysis, where the influence of the number of choice sets on the attribute importance is observed in both designs.

5.1 Sample

The total amount of people who participated in the survey is 195 respondents. Where, in the forced design 100 participants answered the survey and 95 people participated in the voluntary design. To test whether the respondents paid full attention during the survey, there is a test question added to the survey. This question reads: “Select ‘totally agree’ here”. All respondents that answered anything else than ‘totally agree’ are selected as noise. Of all 195 respondents 29% (56 respondents) of them answered this test question wrong, which indicates these respondents did not pay attention. Therefore, these respondents are excluded from the dataset. In the forced design, 31 respondents are excluded and in the voluntary 25. Hence by excluding these respondents the data only provides 139 respondents who were paying attention to have reliable results. The test question is deleted from the data set, no further analysis will be done with this item.

(25)

24 have the lowest level of income, under €20.000,-, since most of them are young and starters in the work field. An overview of these descriptions are presented in table 4.

To check whether respondents in the forced and voluntary design are comparable in terms of gender, age, education, employment status, income and household size, a Chi- Square test is provided (Hair, et al., 2009). The results from this chi- squares test stated that the two groups based on these variables, do not significantly differ from each other. The Pearson Chi- Square for these variables are in sequence, age .415, gender .648, education .845, employment status .470, income .468 and finally the Chi- Square for household size is .654. Concluding out of these statistics, the voluntary and forced group are not significantly different from each other, based on these variables and can be compared.

In table 5, the descriptive information, for gender and age, are presented for both designs; small differences are detected. Additional descriptive information are shown in appendix B. Other differences for both designs, regarding the number of answered choice sets, are shown in this table. In the voluntary design, 21 respondents stopped and skipped towards the end of the survey immediately when the survey allowed them to, after choice set number 12. 12 Respondents finished the survey between choice set 13-22. Surprisingly, 36 of the respondents finished all 24 choice sets in the voluntary design. Finally, in the forced design 57 respondents answered all 24 choice sets and 13 respondents stopped before completing 24 choice sets, of which 7 stopped after the 12th_{choice set.}

Table 4 Sample Data description

Number of participants 139 (excluded respondents 28.72%)

Design 70 forced 69 voluntary

Gender Female 78.4% Male 21.6%

Age <18 year 19.4% 19-29 year 54.0% >30 year 26.6%

Education Primary or high

school 33.8% Vocational education 30.9% Bachelor degree 26.6%

Master and other 8.7% Employment status Student

(26)

25 Table 5

Descriptive information of the voluntary and forced design

Design Forced (70 respondents) Voluntary (69 respondents) Gender female 80.0 % / male 20.0% female 76.8 % / male 23.2 % Age < 18 year 19-29 year 30-39 year > 40 year 20.0% 57.1% 15.7 % 7.1 % < 18 year 19-29 year 30-39 year > 40 year 18.8 % 50.7 % 15.9 % 14.5 % Answered choice sets 12 choice sets 13 choice sets 14 - 22 choice sets 24 choice sets 7 respondents 5 respondents 1 respondent 57 respondents 12 choice sets: 13 choice sets: 14 - 22 choice sets: 24 choice sets: 21 respondents 4 respondents 8 respondents 36 respondents 5.2 Factor analysis

The four variables that drives the behavior in the PWYW model; fairness, reciprocity, self- images and norm conformity, are measured by 13 different items. Factor analysis (exploratory and confirmatory) and structural equation modeling (SEM) are statistical techniques that can be used to decrease the amount of variables into a smaller number of variables by examining the covariation between the variables (Schreiber, et al., 2006). According to Schreiber, et al., (2006), the number of items needs to be reduced, since there are 13 independent variables which measure 4 constructs. Two methods are used in this analysis. First, the items are extracted all together and second, the items are extract separately to avoid possible loadings on each other.

5.2.1 Model fit

First, the different items will be extracted at the same time in the factor analysis. Different items measure the same construct, therefore the dimensionality of these items needs to be reduced. The first question arises: can a factor analysis be performed?. To determine whether a factor analysis is appropriate, the sample adequacy has to be > .05 in the KMO (Kaiser- Meyer- Olkin) and Bartlett’s test (Hair, et al., 2009). The sample adequacy determines appropriateness and is based on correlation and partial correlation. The KMO measure of

sampling adequacy is .614 (> .05) shown in table 6. Moreover the model is significant (p < .05). Concluding, the variables are correlated and a factor analysis can be performed.

Table 6

KMO and Bartlett's Test KMO Measure of

Sampling Adequacy

Bartlett's Test of

(27)

26 5.2.2 Exploratory factor analysis

The Principal Component Analysis (PCA) is the most common factor model used (Widaman ,1993). It shows eigenvalues, which indicate the amount of variance explained by each factor. In table 7, component 1, 2, 3 and 4 are selected as 4

new factors. Together they explain 61.86% of the variance, which has to be at least over 60%. Hence these four components explain the majority of the variance within this data. Each of these factors also explain more than 5% of the individual variance which is a prerequisite as well. Finally they all four have an Eigen value > 1.0, which is an requirement for factors (Hair, et al., 2009).

From here, the rotated factor analysis is used. Rotation prevents all variables to load on one factor and minimizes the number of variables which have high

loadings on each given factor. The Varimax procedure is used. In table 8 the rotated component matrix is shown, with the items and their loadings within the factor. The item Fairness 2 (table 3) is deleted, since the loading was too low (< .5). Therefore, further analysis is continued, using these strongly loaded items in the resulting four factors.

Table 8

Rotated Component Matrix

Table 7

Principal Component Analysis Component Initial Eigenvalues

(28)

27 These new factors are formed and named in succession; reciprocity is factor 1, self- image is factor 2, norm conformity is factor 3 and the fourth factor is fairness. These new variables will be used in the OLS regression, in order to test whether these drivers have an influence on the number of choice sets answered.

5.2.3 Confirmatory factor analysis

Furthermore, to check whether there are different outcomes in the upcoming regression analysis, the items of the drivers of motivation are also extracted separately from each other. In the first part of analysis all items were extracted together, therefore, the items per construct are extracted together and not all items together. This results in a small difference in loadings in the resulting factors, these differences are minimal. In this confirmatory method, items mutually do not load on each other, which results in higher loadings within the factor. The KMO measure of sampling adequacy is for all four constructs acceptable. Moreover, the four separate models are significant (p < .05). Thus, it can be stated that the variables are correlated and that a factor analysis can be performed for all four constructs separately. Results show that separately, they explain > 60% of the variance. Hence, these four components explain the majority of the variance within this data. Each of these factors also explain over 5% of theindividual variance, which is acceptable. Finally, they all four have an Eigen value > 1.0. The item F2 (fairness) is similar as in the first method, deleted in this method, due to low loading in the factor. All steps are the same, however they have a small difference in the factor loadings. The factors from this second method will be used in the following estimates.

5.2.4 Internal consistency

Before the four factors will be used to run the model of Ordinary Least Squares Regression in the next paragraph, internal consistency have to be determined. Therefore, the Cronbach’s Alpha will be provided (Hair, et al., 2009). This measures the strength to proceed with these dimensions instead of the original items. First, the Cronbach’s alpha for the factor reciprocity with all four items is .743 (α > .6) which is a good outcome. Second, the Cronbach’s

alpha for the factor self- image, with all three items, is .620 (α > .6), which is also acceptable. For the third factor, norm- conformity, the Cronbach’s alpha with two items is .600 (α = .6), which is not high however acceptable. Finally, for the factor fairness the Cronbach’s alpha with the three items is .482 which is (α < .6) too low. After the removal of item F2, also item F1 is

removed (see table 3), the results show a α= .556. This is still too low, since it is < .6. Therefore, the Table 9

Internal consistencies

Factor Cronbach’s alpha

Reciprocity .743

Self- image .620

Norm- conformity .600

(29)

28 factor fairness has a low internal consistency with two and three items. Since this is one of the four drivers of voluntary behavior of respondents, this variable will still be included in the OLS with three items.

5.3 Effect drivers on number of choice sets answered

The estimation of the regression model is described in two parts. First, the model fit is explained, then the parameter estimates are determined. This analysis answers the research question and hence the hypothesis. The number of choice sets answered is the dependent variable and the drivers; reciprocity, self- image, norm conformity and fairness, the independent variables. This analysis is dummy coded for the different conditions, forced (0) and voluntary (1).

5.3.1 Model fit

First, the model fit can be evaluated by checking the 𝑅2 and the adjusted 𝑅2 (Keller, 2008). The 𝑅2 and the adjusted 𝑅2 are both acceptable, with .163 respectively .141. Hence the overall model is significant, with a p- value <.05. This means that 14% of

the variance in the dependent variable is explained by the independent variables. Second, the estimated regression model is overall significant, this can be seen by the statistic and the p-value. The model has a F-statistic of 7.372 and is significant (p < .05). This shows that this model explains enough variance. In table 10 these statistics are shown.

5.3.2 Parameter estimates

In table 11 the parameter estimates are shown. Looking at the constant 𝛽0 with a β of 22.061 (p =

.000) shows that participants answer 22 choice sets in the forced situation. These respondents are supposed to answer 24 choice sets in the forced design. However, this forced design consist of 57 respondents who answered 24 choice sets, and 13 who did not complete the survey and quitted between 12 and 24 choice sets. Observing the dummy, which are respondents who participated in the voluntary design, they answered 3.399 questions less than the forced design participants (β = -3.399) (p = .000). They stopped on average earlier, they answered 19 choice sets. The motivational driver norm conformity shows a positive influence on the number of choice sets answered (β = 1.063, p = .003). Meaning, for a one scale unit increase in norm conformity the model predicts that the number of choice sets will increase by 1.063, holding all of the independent variables fixed. From table 11, it can be concluded from the standardized beta that the driver norm conformity is the strongest driver, then reciprocity, self- image and lastly fairness with standardized

(30)

29 betas of .201 respectively, .100, .098 and .059. However, the drivers reciprocity, self- image and fairness do not have significant effect on the number of choice sets answered, since the p > .05. As can be seen in the table, the model is checked on multicollinearity. The VIF statistics in this estimated model are not exceeding 1.025, therefore there is no multicollinearity. Meaning that the variables are not strongly correlating with each other (Hair, et al., 2009).

Table 11 Parameter estimates Unstandardized Standardized t P- value Collinearity statistics Beta Std. Error

Beta _Tolerance _VIF

Forced 22.061 .493 44.728 .000 Dummy Voluntary -3.399 .711 -.322 -4.780 .000 .976 1.025 Reciprocity .529 .353 .100 1.499 .136 .994 1.006 Self- image .517 .355 .098 1.459 .146 .986 1.015 Norm conformity 1.063 .353 .201 3.012 .003 .995 1.005 Fairness .313 .352 .059 .889 .375 1.000 1.000 5.4 Conjoint analysis

In order to generate an answer on hypotheses two and three, a conjoint analysis is conducted. The program Latent Gold is used to generate the outcomes of the conjoint analysis. Within this program the attribute importance will be measured. Investigated is whether the length of the survey has an influence on the attribute importance. Additionally, the effect of the forced and the voluntary designs on this attribute importance are measured. For the conjoint analysis the dataset has to be reduced to participants who only answered all 24 choice sets, since the attribute importance is measured and compared for both conditions. Therefore, all participants in both designs who did not complete 24 questions, are deleted from the dataset and not taken into further analysis. Resulting in a voluntary design of 36 respondents and 57 respondents in the forced design.

5.4.1 Estimation of the model

(31)

30 In the first step the variables are generated in Latent Gold, all five attributes are labeled nominal in both datasets (1-12 and 13-24). Both the choice sets 1-12 and 13-24 are observed with the two classes forced (1) and voluntary (2). This step is taken, to observe whether the variables can be taken into further analysis. Checking the Wald statistics shows whether all variables can be used in the model. As can be seen in table 12, for choice set 1-12, the attributes; airline, number of stops and allowed weight baggage, are not significant, since p > .05. However, the variables destination and ticket price are significant, with p = .012 and .014 (p < .05). The right columns presents the estimates for choice set 13-24. Concluding, all attributes and the Wald statistic show significance, except the attribute airline (Wald = 5.7508, p > .05). These insignificant attributes are necessary to calculate the attribute importance and to compare both conditions in the subsequent phases of the conjoint experiment. Therefore, these attributes are still included in the model knowing their insignificance.

After these estimates, graphs are drawn with all attributes as nominal selected first. The utilities of all attributes are observed in the graphs to check if there could be a linear relationship. When graphs show possible linear relationships, different models are going to be estimated in Latent Gold. Whereby the attributes change from nominal to numeric when the graphs might look linear. The goodness of fit criteria for the different models, presented in table 13 and 14, show whether the attributes has to be used as linear or nominal. The attributes airline and destination are nominal, since these are brands or choices, these are not numeric attributes. The other three attributes are labelled and observed as numeric, which is described below.

Table 12

Estimates of conjoint analysis (all attributes nominal) Choice set 1-12 Choice set 13-24

Attributes Wald p- value Wald p- value

Airline .7155 .70 5.7508 .22

Destination 8.8783 .012 171.7027 4.5e-36

Number of stops 4.6796 .096 176.0062 5.4e-37

Allowed weight baggage .1961 .91 15.7710 .0033

(32)

31 The graphs of possible linear relationships are shown in figure 3. By drawing those graphs it gets more clear whether an attribute could have a linear relationship (straight increasing or decreasing line). Concluding out of the graphs, all three the attributes; amount of stops, ticket price and allowed weight of baggage, might have linear relationships in both data sets (1-12 and 13-24). The utility graph of the attribute amount of stops seems to be most linear of the three, due to the near straight line. Therefore, the model is estimated again, with these three attributes separately as being linear and jointly, while the rest of the attribute stay part- worth. This results in eight different models. The resulting models and the model fit criteria are presented in table 13 and 14 for the choice set 1-12 and 13-24.

Figure 3

(33)

32 Table 13

Goodness of fit criteria (choice sets 1-12) LL (β*) LL(0) Pseudo 𝑹𝟐 𝑹𝟐 adjust Nr para- meters

Hit rate Chi- square BIC CAIC(LL) Model 1 -1022.76 -2452.10 .5829 .5743 21 58.60% 2858.69 2140.70 2161.70 Model 2 -1024.41 -2452.10 .5822 .5745 19 58.60% 2855.38 2134.94 2153.94 Model 3 -1073.89 -2452.10 .5621 .5543 19 55.91% 2756.43 2233.89 2252.89 Model 4 -1023.42 -2452.10 .5826 .5749 19 58.87% 2857.36 2132.97 2151.97 Model 5 -1074.97 -2452.10 .5616 .5547 17 55.82% 2754.27 2226.99 2243.99 Model 6 -1025.12 -2452.10 .5819 .5750 17 58.33% 2853.97 2127.29 2144.29 Model 7 -1074.65 -2452.10 .5617 .5548 17 55.82% 2754.90 2226.36 2243.36 Model 8 -1075.78 -2452.10 .5613 .5552 15 55.47% 2752.65 2219.55 2234.55

Model 1: All five attributes part-worth

Model 2: Attribute amount of stops linear, rest part-worth Model 3: Attribute price linear, rest part-worth

Model 4: Attribute baggage linear, rest part- worth

Model 5: Attributes amount of stops and price linear, rest part- worth Model 6: Attributes amount of stops and baggage linear, rest part- worth Model 7: Attributes price and baggage linear, rest part- worth

Model 8: Attributes amount of stops, price, and baggage linear, rest part- worth

Table 14

Goodness of fit criteria (choice sets 13-24) LL (β*) LL(0) Pseudo 𝑹𝟐 𝑹 𝟐 adjust Nr para- meters

Hit rate Chi- square BIC CAIC(LL) Model 1 -1046.63 -2452.10 .5732 .5646 21 58.24% 2810.95 2140.70 2209.44 Model 2 -1047.04 -2452.10 .5730 .5653 19 58.33% 2810.12 2134.94 219921 Model 3 -1094.14 -2452.10 .5538 .5460 19 53.41% 2715.93 2233.89 2293.39 Model 4 -1047.45 -2452.10 .5728 .5652 19 58.07% 2809.30 2132.97 2200.02 Model 5 -1094.22 -2452.10 .5538 .5468 17 53.41% 2715.76 2226.99 2282.50 Model 6 -1047.86 -2452.10 .5727 .5657 17 58.07% 2808.49 2127.29 2189.77 Model 7 -1094.95 -2452.10 .5535 .5465 17 53.23% 2714.31 2226.36 2283.95 Model 8 -1095.02 -2452.10 .5534 .5473 15 53.49% 2714.16 2219.55 2273.03

Model 1: All five attributes part-worth

Model 2: Attribute amount of stops linear, rest part-worth Model 3: Attribute price linear, rest part-worth

Model 4: Attribute baggage linear, rest part- worth

Model 5: Attributes amount of stops and price linear, rest part- worth Model 6: Attributes amount of stops and baggage linear, rest part- worth Model 7: Attributes price and baggage linear, rest part- worth

(34)

33 The following step in this analysis is to check which of the models fits the best. For both the data set situations the different information criteria displayed in both tables will be carefully observed. First of all, the 𝑅2adjusted and the hit rate are observed. These criteria are the critical measurements for selecting the best model. Based on these criteria the model that fits the best will be chosen for further calculations for the attribute importance. First, the criteria of choice sets 1-12 is observed. The hit rate model is high in the models 1, 2, 4 and 6. With the highest hit rate of 58.87% in model 4 and the second highest hit rate of 58.60% in both models 1 and 2. The hit rate gives percentages, these explain the percentage of observed choices that are predicted correctly for this model. Second important criteria is the 𝑅2 adjusted. The 𝑅2 adjusted in model 4 and 6 are highest for choice set 1-12, namely .5749 and .5750. However, the 𝑅2 adjusted in model 2 is also high with .5745. These criteria are also observed for the other choice sets.

In the choice sets 13-24, the highest hit rate is observed in model 2, and subsequently in model 1, with 58.33% and 58.24%. Meaning that 58.33% of the observed choices are correctly predicted by the model. Hence, repeatedly, in the models 4 and 6, the hit rate is high, with both a hit rate of 58.07%. When observing the 𝑅2 adjusted in the resulting models, the highest is observed in model 6, with .5657. Followed by model 2 with .5653 and model 4 with .5652.

From these two tables, it can be concluded that the models 1, 2, 4 and 6 do look like each other according these described criteria. The model that is selected to take into further analysis is model 2. In both situations the hit rate and the 𝑅2 adjusted are both high. A lot of the variance is explained in this model 2. Differences with the models 4 and 6 are very small. The utility graph in figure 3 showed the most linear relationship line for the attribute the amount of stops. Therefore, the model 2 is selected. This model has all part- worth attributes, except for the amount of stops, which is linear. 5.4.2 Relative importance

(35)

34 lowest attribute level. In the linear attribute the amount of stops ‘0 stops’ and ‘2 stops’ are multiplied by the utility -1.0933 (forced) and -1.3224 (voluntary). In table 15 the attribute importance for all attributes are calculated for both designs (forced and voluntary) separately. Hence the differences in importance between these two are calculated in the third and fourth column. In appendix C the calculations of the relative attribute importance are shown in two tables.

As can be seen in this table, the attribute amount of stops is by far the most important attribute in both choice sets, with a relative importance of 48.51% (forced) and 43.82% (voluntary) for choice set 1-12 and 53.98% (forced) and 40.62% (voluntary) for the choice set 13-24. After that the most important attribute for respondents were destination and ticket price. First, the influence of the number of choice sets on the attribute importance is investigated, this is investigated in the forced design. Second, the influence of forcing respondents to finish the survey on the attribute importance is investigated. Therefore, the attribute importance is both, the voluntary and the forced, design are compared.

For the amount of stops the difference in importance between a short survey and a long survey is large. Especially in the forced design this importance changes up to 5.47% more important than when asked in the beginning of the survey. This is a remarkable change in importance. Other changes in attribute importance for the forced design is the decrease in importance for the attribute, airline

Table 15

Relative Attribute Importance and difference in importance between choice sets

Attribute Importance 1-12 Importance 13-24 Difference in importance Absolute difference Airline 4.87 % (forced) 6.00 % (voluntary) 1.05 % (forced) 6.16 % (voluntary) - 3.82 % .16 % 3.82 % .16 % Destination 21.18 % (forced) 23.93 % (voluntary) 22.29 % (forced) 22.78 % (voluntary) 1.11 % - 1.15 % 1.11 % 1.15 % Amount of stops 48.51 % (forced)

43.82 % (voluntary) 53.98 % (forced) 40.62 % (voluntary) 5.47 % - 3.20 % 5.47 % 3.20% Allowed baggage 5.89 % (forced)

3.85 % (voluntary) 7.39 % (forced) 4.97 % (voluntary) 1.50 % 1.12 % 1.50 % 1.12 % Ticket price 19.55 % (forced)

(36)

35 when the survey is longer, with a decline of 3.82%. Finally and surprisingly, the attribute ticket price does not increase in importance like literature concluded, however it decreased with 4.25% for the forced design.

For the voluntary design, the attribute importance for ticket price increases during the survey. In the beginning of the survey the importance is 22.39% and during the second part of the survey the importance increases with 3.08%, up to 25.47%. Contrary to the forced design the attribute importance for the amount of stops decreases with 3.20%. The attribute, airline barely changes, with .16%. Therefore, in both designs, there are different changes in attribute importance. Finally in both designs the attributes, destination and allowed weight baggage do not have a remarkable change in attribute importance due to the length of the survey.

Comparing the two designs with each other results in that changes in attribute importance during the survey for the voluntary design are fewer than in the forced design for four of the five attributes. For the attributes airline, amount of stops, allowed baggage, and ticket price, the differences in attribute importance in the shorter survey, compared to the longer survey are smaller in the voluntary design. It is only for the attribute destination that the difference is greater in the voluntary design compared to the forced design, with a small difference of .04%, which is almost negligible. To measure the differences in attribute importance between the two designs, the average differences are measured. This is calculated by summing up the absolute differences in attribute importance and dividing this by 5 (attributes) for both designs. For the forced design, the average difference is 16.15 divided by 5 attributes, which is 3.23. For the voluntary design, this is 8.71 divided by 5 attributes, which is 1.742.

6. Discussion

Key findings of previous literature are discussed in this chapter together with the findings of this study. This chapter presents all seven hypotheses with their outcomes.

6.1 Discussion

The main goal of this research is to examine the effect of a voluntary design on the number of choice sets, and whether there are differences in attribute importance when using a voluntary design as compared to forcing respondents.

6.1.1 The effect of voluntary design on the number of choice sets answered

The most optimal amount of choice sets in a conjoint analysis and the analysis of the drivers of willingness to respond

The most optimal amount of choice sets in a conjoint analysis

and the analysis of the drivers of willingness to respond

What is the influence of forcing respondents on their choice behavior?

By

The most optimal amount of choice sets in a conjoint analysis

and the analysis of the drivers of willingness to respond

What is the influence of forcing respondents on their choice behavior?

Abstract

Preface

Table of Contents

1. Introduction

2. Literature review

3. Conceptual model

4. Research methodology

𝛽

𝑦

= 𝛽

+ 𝛽

𝑋

+ 𝛽

𝑋

+ 𝛽

𝑋

+ 𝛽

𝑋

+ ℇ

Where,

𝛽

𝑈

= 𝑉

+ 𝜀

𝑉

= ∑

𝛽

𝑥

𝑉

𝑉

= 𝛽

𝐴𝐼𝑅

+ 𝛽

𝐷𝑒𝑠𝑡

+ 𝛽

𝑆𝑇𝑂𝑃

+ 𝛽

𝐵𝐴𝐺

+ 𝛽

𝑃𝑅𝐼𝐶𝐸

𝛽

𝛽

𝛽

𝛽

=

5. Results

6. Discussion