Text vs Image Presentation in a Choice Based Experiment on Willingness to Respond, Accuracy, Willingness to Pay and Reliability

Text vs Image Presentation in a Choice Based Experiment on

Willingness to Respond, Accuracy, Willingness to Pay and

Reliability

By:

Niels Wesselink

Abstract

We describe how images versus text in an experimental design of a choice set affect the accuracy (relative attribute importance), willingness to pay (WTP), willingness to response (WTR) and reliability. Also, we test the moderating and direct effect of WTR on the relationships of images versus text on accuracy, WTP and reliability. The study is based on a burger study. Hypotheses were tested by a randomized in-between subjects sample design in a multinomial logit framework. The most important findings are that the image design did affect the accuracy, increased WTP, lowered reliability and increased the WTR. Also, a text condition in combination with a high WTR lead to higher reliability. At last there are no differences in price sensitivity or WTP due to a higher or lower WTR.

Keywords: Choice Experiments, Conjoint Analysis, Fatigue, Learning, Number of Choice Sets, Choice Set Design, WTR

Supervised by: dr. F. (Felix) Eggers, Assistant Professor University of Groningen

3

1. Introduction

Every day companies are trying to conquer the market with new products. One of the ways to identify what consumers really like in the market and to see whether a product meets the demands of the market is preference measurement. Conjoint analysis is one of the most important topics in preference measurement. It helps marketing- and management decision makers by revealing what customers like and prefer (Eggers and Sattler 2011). Also, it can help to reveal preferences in other fields of study, for example healthcare (Ryan and Farrar 2000). In a choice-based conjoint experiment, individuals are presented with a series of choice sets that contain alternatives that differ in attributes and attribute levels and are asked to choose one alternative that they prefer most. When price is one of the attributes consumers’ willingness to pay for each attribute can be computed from estimates of econometric models (Chung, Boyer, and Han 2010).

Different aspects of the conjoint analysis can be researched, for example the effects of craft in conjoint analysis i.e. to what extend different aspects of conjoint analysis affect the reliability and accuracy of the results. Hauser, Eggers, and Selove (2016) identify that using pictures instead of text-based descriptions increase accuracy and reliability. They also show that using incentive alignment increases reliability. Another paper by Ding et al. (2005) backs up this finding, by stating that salient incentives lead to higher price sensitivity and higher reliability. Besides this research, reliability can also be increased by increasing the number of choice sets i.e. collecting more data (Johnson and Orme 2002; Eggers 2015). On the other hand, there is also much contradicting research that state that more choice sets answered lead to lower accuracy and reliability (DeShazo and Fermo 2002; Caussade, Ortúzar, Rizzi and Hensher 2005). The main reason for this is the cognitive burden that is associated with answering more choice sets and therefore deplete resources. Many researchers looked at the optimal number of choice sets, for example Chung et al. (2010) argues that the optimal should be 6 sets, while Caussade et al. (2005) and Bradley and Daly (1994) say that the optimal should be around 10.

(Eggers 2015). The choice set designs are based on text or images and mostly presented point by point instead of a description. Besides, the contexts’ differ substantially, (Caussade et al. 2005) base their findings on a transport context, while (Carlsson, Mørkbak and Olsen 2012) use chicken breast filet as their choice experiment. Considering the differences, it is highly likely and logical that the studies come up with mixed results.

5

For marketing researchers, this paper can add value to the way they should design surveys while doing conjoint analyses.

The structure of this paper is as follows. First, we will start with the literature review in which we state our constructs, hypotheses and research design. Next, the methodology will be explained and the variable operationalization will be done. Thereafter, the data obtained will be described. After that, there will be a section where the analysis together with the results are presented. At last there will be a discussion and conclusion section which will include managerial implications and limitations to this research, also we will try to advice on implications for further research.

2.

Literature review

Text vs. Image

Stimuli in conjoint experiments can be described, among others, via text of images. Text or a verbal is defined as a series of words to describe the attribute levels in a choice set. An image or visual we define as being a picture that presents the attribute levels in a choice set, so without any words nor description. The main theoretical difference between the processing of text versus image lays in the system 1 and 2 thinking. System 1 thinking is fast and intuitive thinking, whereas system 2 is more deliberative reasoning (Pocheptsova et al. 2009). We would like to know whether there are differences in processing and whether it will lead to differences in accuracy, WTP, reliability, and WTR.

2.1.1 Effect of Text versus Images on Accuracy

Accuracy is defined as the true relative importance of attributes. Because the true relative importance cannot be observed, we can state that accuracy is affected when we see a change in relative attribute importance (Czajkowski, Giergiczny and Greene 2014). Table 1 gives an overview of studies that looked at the differences of image versus text on accuracy. Anderson

attribute importance in the image condition compared to the text condition. On the other hand, Holbrook & Moore (1981) and Domzal and Unger (1985) did not find any difference between image and text in change in relative attribute importance. According to Vriens et al. (1998) judgements easier made in the image condition which lowers fatigue and therefore we believe there will be a change in attribute importance between text and image, lead to the following hypothesis:

H1a: Images, compared to text, does have an effect on the accuracy of a choice based experiment

Table 1. Review of studies investigating image versus text on accuracy

Anderson (1987) Domzal and Unger (1985) Holbrok and Moore (1981) Louviere, J. J. et al. (1987) Smead et al. (1981) Vriens et al. (1998) Change in attribute

importance (Accuracy) Yes No No Yes Yes Yes

Category of study Clothing Watches Clothing Skate parks Coffee Car stereo's

Number of attributes 3 3 5 3 9 7 Design Within-subjects Between-subjects Between-subjects Between-subjects Between-subjects Between-subjects Heterogeneous/homogenous

respondents Homogenous Heterogeneous Homogenous Heterogeneous Heterogeneous Homogenous

2.1.2

Effect of Text versus Images on WTP

7

there is a higher WTP and lower price sensitivity in the image condition, leading to the following hypothesis:

H1b: There is a lower price sensitivity and a higher WTP for images compared to the text based choice set

.

2.1.3 Effect of Text versus Images on Reliability

Reliability we define as overall consistency of a measure. An experiment should be repeated a significant number of times and shows the same results each time DeShazo and Fermo (2002). According to Dellaert, Donkers and Soest (2012); Haaijer, Kamakura and Wedel (2000); Otter, Allenby and van Zandt (2008) the largest threat to reliability in a choice based experiment is resource depletion or fatigue, therefore lower consideration times would lead to significant more reliable decision making. Pocheptsova, Amir, Dhar and Baumeister (2009) take the psychological perspective and look at system 1- and 2 thinking. They argue that system 1 thinking, which is the case with interpreting a visual, leads to less depletion than system 2 (for verbal) thinking that is more effortful. Also, Pieters and Warlop (1999) add to this by stating that text is filtered more compared to pictorial information in choice tasks, which results in less reliable answers. So, people rely on more simple and intuitive decision making once their resources had been depleted, leading to biased decisions. Where most conjoint analyses are limited to a certain amount of choice sets we argue that one should best use images instead of text in order to limit fatigue, leading to more reliable decision making. This leads to the following hypothesis:

2.2 Effect of Text versus Images on WTR

The number of choice sets answered or willingness to respond (WTR) is defined as the total of choice sets successfully completed and can have a value of any positive number or zero (Eggers 2015). According to Eggers (2015) fatigue has the largest negative effect on WTR. In order to limit fatigue, it is important to look at how we can present choice sets in such way that resources deplete slowest. Levie and Lentz (1982) summarizes 55 experiments that research text versus image on learning. They conclude that images lead to higher retention of subjects and to greater attention versus text. According to Jensen (2008) eyes are capable of registering 36000 visuals per hour and between 80- and 90% of all information is visual. A study by Potter, Wyble, Hagmann and McCourt (2014) shows that it takes only 13ms to understand the meaning of a picture, while Holcomb and Grainger (2006) state that the processing time of a word takes twice as long, which means that the effort one has to take to interpret a picture is less than with a text. At last Townsend and Kahn (2014) conclude that respondents prefer visuals rather than verbal depiction of stimuli in choice sets. Therefore, we propose that presenting choice tasks as images instead of text decrease fatigue and therefore will increase the number of choice sets answered, leading to the following hypothesis:

1

2.3 Direct- and Moderating Effects of Number of Choice Sets

2.3.1 Accuracy

Table 2 provides a list of studies that have researched the effect of the number of choice sets on the accuracy, WTP and reliability. Regarding accuracy,Carlsson et al. (2012) found that more choice sets answered resulted in lower price sensitivity, however they did not find any change for other attributes. On the other hand Bech et al. (2011); Czajkowski et al. (2014); Day et al. (2012); DeShazo and Fermo (2002) did find differences in attribute importance when answering more choice sets. Bech et al. (2011) even found that with 17 choice sets respondents said they based their choice only on one attribute while in the 5 choice condition respondents based their decision on almost all attributes presented. Czajkowski et al. (2014) stated that the change in attribute importance was due to the learning effect. Based on this literature we assume that accuracy changes when respondents answer more choice sets, leading to the following hypothesis:

H3a: The number of choice sets lead to different relative attribute importance’s.

The higher WTR, the more resources deplete. In combination with a text condition, respondents will fatigue even sooner, leading to different attribute importance’s compared to the image condition, leading to the following hypothesis:

H4a:

2.3.2 WTP

and Orme (2002) conclude that as the number of choice sets answered increase the price attribute becomes more important, so higher price sensitivity. Considering these studies, we believe that a higher WTR lead to a strategy change where respondents mainly focus on one attribute. Therefore, we believe that more choice sets lead to higher price sensitivity and higher WTP, leading to the following hypothesis:

H3b: The higher the WTR the higher price sensitivity and willingness to pay.

For the image condition fatigue is lower than the text condition. Including the effect of answering more choice sets to the image condition will lead to depletion of resources sooner and therefore lead to a higher price sensitivity and a lower WTP then with lower choice sets. Therefore, images and a higher WTR will lead to lower WTP and higher price sensitivity, leading to the following hypothesis:

H4b: A higher WTR lowers the WTP and increases the price sensitivity in the image condition (and vice versa), so has a negative moderating effect.

2.3.3 Reliability

13

H3c: The number of choice sets answered show an inverted U-shape pattern, first a positive effect, but later a negative effect on reliability.

Concluding from 2.1.3 the fatigue effect is the main threat to reliability, therefore a higher WTR in combination with the text condition would lead to lower reliability then a low WTR in combination with text. Therefore, we believe that WTR has a negative moderating effect on the relationship of text versus image on reliability leading to the following hypothesis:

H4c: A higher WTR in both the text and condition lowers reliability versus a lower WTR in the text and image condition, therefore WTR has a negative moderating effect on the relationship of text versus image on reliability.

2.4 Conceptual Framework

IV’s

Text vs. Images

DV’s

a. Accuracy

b. WTP

c. Reliability

Moderator

Number of Choice Sets

Answered (WTR

)

H1a, H1b, H1c

H3a, H3b, H3c H2

H4a, H4b, H4c

3. Methodology

3.1 Experimental design

The empirical study applied voluntary choice sets to a randomized between subject’s design. The study was conducted among students of the master marketing course digital marketing intelligence of the University of Groningen. The other respondents are friends and family of the three master thesis students under supervision of F. Eggers at the University of Groningen. The research context was fast food, in particular a burger study with a maximum of seven attributes: (1) lettuce, (2) onions, (3) tomatoes, (4) meat type, (5) cheese slices, (6) sauce and (7) price of the burger. Attribute levels ranged from 2 to 4, see table 3 for an overview of the attributes with their levels. The number of alternatives was kept constant (3) for every choice set and consisted of a combination of three burgers with a variation of attributes and levels. There was minimal overlap between the choice sets. The survey consisted of 57 tasks in total, after the fourth set respondents were informed that they could skip to the end of the survey at any time, using a link on the bottom of each page. In both conditions respondents were asked to indicate which burger they preferred most, after that they were asked to indicate (yes/no) whether they realistically would purchase their burger choice.As a reward for participation, respondents were able to take part in a lottery in the survey, we offered the opportunity to win 5 x €10. The respondent that is lucky received the burger that is closest to their preference for the price indicated, if the price is lower than €10 they will receive the burger + the difference in cash. If the burger was not chosen, they would receive €10. Also, for the respondents that took part in the Digital Marketing Intelligence class there was a 0,3 extra point bonus on their assignment. Incentive alignment was not one of the conditions, all respondents received this condition. Respondents were randomized among the two conditions with either the text or image, see figure 2a and 2a for examples of both conditions.

Table 3. Attributes and Attribute Levels

Attributes Levels

Lettuce Yes - No

Onions Yes - No

Tomatoes Yes - No

Meat Type Beef - Chicken

Meat Patties 1 - 2

Cheese Slices 0 - 1 - 2

Sauce Ketchup - Barbecue – Chili - Mayonnaise

15

3.2 Measures

The conceptual model in figure 1 can be estimated by a standard multinomial logit model. We

will analyze choices of consumers of different burger combinations. The theoretical framework

for choice behavior is the random utility theory, which states that the overall utility U of

individual i for an object j is a latent construct that includes a systematic component V and an

error component , which catches all effects that are not in the model (Manski 1977):

(1)

According to McFadden (1974), consumers implicitly compare the utility of the choice options

and choose the alternative that gives them the highest utility. Where the chosen option can be

any alternative i chosen from choice set J = {1,…,57}:

(2)

𝑝𝑟𝑜𝑏(𝑖|𝐽) =

𝑗=1 Figure 2a. Example of a choice set in the image

condition

3.2.1 Accuracy

Accuracy; for the relative attribute importance, we will estimate the standard multinomial logit model with price effect coded (Eq. 3)2_{. We will calculate the ranges for each attribute (Eq. 4), after that we will} take the sum and divide the ranges by the total in order to obtain the relative attribute importance’s (Orme, 2009 Ch.9). We will do this for the text as well as the image condition. For the WTR on accuracy we will use groups; <5, >5 - <16 and >15 sets. We will obtain relative attribute importance’s for each group, in comparing results we can see whether there are direct or moderating effects visible.

(3) 𝑉𝑖= 𝛽𝑀𝑀𝑒𝑎𝑡𝑖+ 𝛽𝐶𝐶ℎ𝑒𝑒𝑠𝑒𝑖+ 𝛽𝐿𝐿𝑒𝑡𝑡𝑢𝑐𝑒𝑖+ 𝛽𝑇𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠𝑖+ 𝛽𝑂𝑂𝑛𝑖𝑜𝑛𝑠𝑖 + 𝛽𝑆𝑆𝑎𝑢𝑐𝑒𝑖+ 𝛽𝑃𝑃𝑟𝑖𝑐𝑒𝑖 + 𝛽_𝑁𝑀𝑁𝑜𝑛𝑒 + 𝐼𝑚𝑎𝑔𝑒 ∙ (𝛽_𝑀𝑀𝑀𝑒𝑎𝑡_𝑖+ 𝛽_𝐶𝑀𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝑀𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖 + 𝛽_𝑇𝑀𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠_𝑖+ 𝛽_𝑂𝑀𝑂𝑛𝑖𝑜𝑛𝑠_𝑖+ 𝛽_𝑆𝑀𝑆𝑎𝑢𝑐𝑒_𝑖+ 𝛽_𝑃𝑀𝑃𝑟𝑖𝑐𝑒_𝑖+ 𝛽_𝑁𝑀𝑁𝑜𝑛𝑒) (4) 𝑎𝑏𝑠(max 𝑙 [𝛽𝑘] − min𝑙 [𝛽𝑘]) 3.2.2 WTP

For WTP we will use a standard multinomial logit model again to estimate the effects of all attributes and levels on the selection dummy (Eq. 3). This time we use price as a linear variable. By dividing the estimates by the absolute price estimate we will obtain the WTP for each attribute level (Orme, 2009 Ch.9). We will do this for text as well as image. In order to test whether WTR has a direct effect we will estimate the base model, but including a direct effect with number of sets answered (Eq. 5). Also, we will estimate a second model with the interaction effect of WTR on images (Eq. 6) in order to find a moderation effect.

17 (5) 𝑉_𝑖= 𝛽_𝑀𝑀𝑒𝑎𝑡_𝑖+ 𝛽_𝐶𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖+ 𝛽_𝑇𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠_𝑖+ 𝛽_𝑂𝑂𝑛𝑖𝑜𝑛𝑠_𝑖+ 𝛽_𝑆𝑆𝑎𝑢𝑐𝑒_𝑖+ 𝛽_𝑃𝑃𝑟𝑖𝑐𝑒 + 𝛽_𝑁𝑀𝑁𝑜𝑛𝑒 + 𝐶ℎ𝑜𝑖𝑐𝑒 𝑆𝑒𝑡𝑠 𝐴𝑛𝑠𝑤𝑒𝑟𝑒𝑑 ∙ (𝛽_𝑀𝐶𝑀𝑒𝑎𝑡_𝑖+ 𝛽_𝐶𝐶𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝐶𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖 + 𝛽𝑇𝐶𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠𝑖+ 𝛽𝑂𝐶𝑂𝑛𝑖𝑜𝑛𝑠𝑖+ 𝛽𝑆𝐶𝑆𝑎𝑢𝑐𝑒𝑖+ 𝛽𝑃𝐶𝑃𝑟𝑖𝑐𝑒 + 𝛽𝑁𝐶𝑁𝑜𝑛𝑒) (6) 𝑉_𝑖 = 𝛽_𝑀𝑀𝑒𝑎𝑡_𝑖+ 𝛽_𝐶𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖+ 𝛽_𝑇𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠_𝑖+ 𝛽_𝑂𝑂𝑛𝑖𝑜𝑛𝑠_𝑖+ 𝛽_𝑆𝑆𝑎𝑢𝑐𝑒_𝑖+ 𝛽_𝑃𝑃𝑟𝑖𝑐𝑒 + 𝛽_𝑁𝑀𝑁𝑜𝑛𝑒 + 𝐼𝑚𝑎𝑔𝑒 ∙ (𝛽_𝑀𝑀𝑀𝑒𝑎𝑡_𝑖+ 𝛽_𝐶𝑀𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝑀𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖+ 𝛽_𝑇𝑀𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠_𝑖+ 𝛽_𝑂𝑀𝑂𝑛𝑖𝑜𝑛𝑠_𝑖 + 𝛽𝑆𝑀𝑆𝑎𝑢𝑐𝑒𝑖+ 𝛽𝑃𝑀𝑃𝑟𝑖𝑐𝑒 + 𝛽𝑁𝑀𝑁𝑜𝑛𝑒) + 𝐶ℎ𝑜𝑖𝑐𝑒 𝑆𝑒𝑡𝑠 𝐴𝑛𝑠𝑤𝑒𝑟𝑒𝑑 ∙ [𝐼𝑚𝑎𝑔𝑒 ∙ (𝛽_𝑀𝑀𝐶𝑀𝑒𝑎𝑡_𝑖 + 𝛽_𝐶𝑀𝐶𝐶ℎ𝑒𝑒𝑠𝑒_𝑖+ 𝛽_𝐿𝑀𝐶𝐿𝑒𝑡𝑡𝑢𝑐𝑒_𝑖+ 𝛽_𝑇𝑀𝐶𝑇𝑜𝑚𝑎𝑡𝑜𝑒𝑠_𝑖+ 𝛽_𝑂𝑀𝐶𝑂𝑛𝑖𝑜𝑛𝑠_𝑖 + 𝛽_𝑆𝑀𝐶𝑆𝑎𝑢𝑐𝑒_𝑖+ 𝛽_𝑃𝑀𝐶𝑃𝑟𝑖𝑐𝑒 + 𝛽_𝑁𝑀𝐶𝑁𝑜𝑛𝑒) 3.2.3 Reliability

In order to test for reliability, we use holdout sets (Orme 2009 and Grover and Vriens 2006). In the survey, we had two holdout sets, choice set 3 and 7 are similar (but alternatives are mixed). We consider a respondent reliable when they chose the same burger in both sets. In order to test this, we will create a cross table on which we will perform a chi-square test, in order to see whether there are differences between text and image. In order to test whether the WTR has a direct effect on reliability and/or a moderating effect we will create three groups3_{: Group 1: <8 choice sets, Group 2: >7 choice sets, but} <16 choice sets and Group 3: >15 choice sets and see whether there are different reliability scores for each group.

3.2.4 WTR

The difference in WTR for text and image will be tested by conducting a Cox proportional hazards survival regression model (Eq. 7)4_{. Additionally, we calculated a simple t-test for comparing means.} Since WTR probably show a skewed result we will also performed a t-test on a log transformed WTR. (7) ℎ_𝑖(𝑡) = ℎ0(𝑡)exp(𝛽1𝐿𝑒𝑐𝑡𝑢𝑟𝑒 + 𝛽2𝐼𝑚𝑎𝑔𝑒𝑠 + 𝛽3𝐺𝑒𝑛𝑑𝑒𝑟 + 𝛽4𝐴𝑔𝑒 + 𝛽5𝑋74𝑎 + 𝛽₆𝑋74𝑏 + 𝛽₇𝑋74𝑐 + 𝛽₈𝑋74𝑑 + 𝛽₉𝑋74𝑒 + 𝛽₁₀𝑋75𝑎 + 𝛽₁₁𝑋75𝑏 + 𝛽₁₂𝑋75𝑐 + 𝛽₁₃𝑋75𝑑 + 𝛽14𝑋76𝑎 + 𝛽15𝑋76𝑏 + 𝛽16𝑋76𝑐 + 𝛽17𝑋76𝑑 + 𝛽18𝑋76𝑓) 4.

Results

In total, 249 respondents completed the survey. Seventy-six of the respondents were part of the Digital Marketing Intelligence class at the University of Groningen, 173 did not take part in the class. The survey was conducted in November 2017. Respondents received randomly the text or image condition, 115 and 112 for image and text respectively. Vegetarians were excluded from the data, due to the fact that we believe this would bias the results, out of the 249, there were 22 vegetarians. The age ranged from 18 to 61, with a median of 24. Also, on average respondents took 606,09 and 621,61 seconds in the image and text condition respectively in order to finish the survey.

4.2 Discrete Choice Experiment Results

Table 4 gives an overview of the model fits of the models estimated. As can be seen the Log Likelihood decreased i.e. became closer to 0 when adding more variables, this means that model (6) is performing best. Also, if we look at the R2 _{and adjusted R}2 _{they are increasing, this is a logical consequence of} adding more variables. Overall, the pseudo R2_{’s are between 0,2 and 0,4 whichis acceptable. On the} other hand, the AIC is lowest for model (3) which indicates this is the best model. At last, all models

19

are significantly better than the NULL-model. The raw output of the Base Model, Model (3), Model (5) and Model (6) can be found in table 17 – 20 in the appendix.

Table 4. Assessment of Choice Experiment Model Fit

Base Model5 _{Model (3) Model (5) Model (6)}

Log Likelihood -2749,1 -2711,3 -2728,3 -2704,6

R2 _{0,2078 0,2187 0,2138 0,2206}

Adjusted R2 _{0,2121 0,2273 0,2051 0,2326}

AIC 5528,228 5482,69 5508 5487,296

P-values Chi-square test 0,00001 0,00001 0,00001 0,00001 NULL-model LL: -3470,42 (227*11,0281124*ln(1/4)

4.2.1 H1a: Image versus text on Accuracy

The results from the images versus text attribute importance’s are in table 5. It turns out that especially for meat type, lettuce, tomatoes and sauce there is rather large difference in attribute importance (-6,35%, 11,14%, 6,85% and -11,88% respectively). So, we can conclude that there is indeed a difference in accuracy between the image and text condition. However, we cannot say whether accuracy is higher or lower in the image- or text condition. Therefore, we can accept H1a stating there are differences in attribute importance between text and image.

Table 5. Attribute importance’s image versus text

Attribute Image Text Difference

Meat type 16,34% 22,68% -6,35% Cheese 13,05% 10,87% 2,18% Lettuce 25,65% 14,51% 11,14% Onions 7,07% 8,32% -1,25% Tomatoes 12,75% 5,90% 6,85% Sauce 6,52% 18,39% -11,88% Price 18,62% 19,32% -0,70%

Mean absolute error 0,058

4.2.2 H1b: Image versus text on WTP

One of the attributes measured in 4.2.2 is WTP, we calculated WTP estimates by dividing the different attribute level estimates over the linear price estimate. We did this by creating the multinomial logit model, where we considered price linear, this resulted into a negative coefficient which is logical. We

did this for the text condition and compared it with the WTP estimates for the image condition. The results are in table 6. The results can be split in two outcomes, the price sensitivity and WTP. First of all, the linear price estimate does not differ significantly between the image and text condition. This result implies that there is no difference in price sensitivity between the conditions. On the other hand, there are significant differences in WTP estimates between text and image. For the image condition, there is a significant higher WTP for lettuce, tomatoes and chili sauce, and a lower WTP for one patty of chicken or beef and barbecue sauce. For example, respondents were willing to pay €1,92 more for lettuce and €1,09 more for tomatoes in the image condition. For the other attributes, there was no significant difference for the WTP estimate. Overall, there is a higher average WTP for the image condition. So, in general we can say that there is no difference in price sensitivity, but there is higher WTP in the image condition. So, we can partly accept our hypothesis since there is on average a higher WTP in the image condition, but price sensitivity is equal.

Table 6. Willingness to Pay Estimates

Attributes Levels Image Text

WTP p-value WTP p-value

Meat type Meat_Chicken_1x -0,52 * 0,04845 -1,03 *** 0,00002

Meat_Chicken_2x -0,20 0,57818 -0,33 0,13320 Meat_Beef_1x -0,72 * 0,01553 0,08 0,70358 Meat_Beef_2x 1,45 0,75000 1,28 *** 0,00010 Cheese Cheese_0x 0,00 1,00000 0,00 1,00000 Cheese_1x 1,76 0,22031 0,99 ** 0,00142 Cheese_2x 1,56 0,63168 1,11 *** 0,00029 Lettuce No_Lettuce 0,00 1,00000 0,00 1,00000 Lettuce 3,41 *** 0,00003 1,49 *** 0,00000 Onions No_Onions 0,00 1,00000 0,00 1,00000 Onions 0,93 0,84615 0,85 *** 0,00066 Tomatoes No_Tomatoes 0,00 1,00000 0,00 1,00000 Tomatoes 1,69 * 0,01080 0,60 * 0,01605 Sauce Sauce.Ketchup 0,01 0,74020 0,10 0,63105 Sauce.Chili 0,20 *** 0,00006 -1,01 *** 0,00002 Sauce.Barbecue -0,54 *** 0,00000 0,87 *** 0,00002 Sauce.Mayonaise 0,33 0,35000 0,04 0,80000 Average WTP 0,55 0,30

Price Absolute estimates 0,38 0,32

21 4.2.3 H1c: Image versus text on Reliability

Out of the 227 respondents 93 answered the first and second holdout and can therefore we used for the reliability analysis. We performed a chi square test and we can conclude that there is a significant effect of image and text on the reliability of the answers (p-value = 0,018), see table 12 in the appendix. From this result, we can conclude that choice sets in the text condition are significantly more reliable than the image condition. The consistent holdouts were 87,50% and 66,04%, text- and image condition respectively. Conclusion; we can reject H1c because reliability is higher in the text condition.

4.2.4 H2: Image versus text on WTR

The averages sets answered was 9,47541 in the text conditions, while the average was 12,520 for the image condition, see table 13. It seems the WTR is higher in the image condition. But, in order to test for this result more formally we conducted a Cox Proportional Hazards Model. It gives clear evidence that respondents that were in the image condition answer more questions, compared to the text condition (p-value = 0,00313). Also, attending the lecture (p-value = 0,01608) and liking to take part in lotteries (p-value = 0,02215) led to a significant higher number of choice sets answered. See table 15 of the appendix for the output. So, according to the Cox Proportional Hazards images lead to a higher WTR, therefore, we can accept H2.

4.2.5 H3a and H4a: WTR on Image versus text on Accuracy

As we have seen in 4.2.1, there are differences between text and images, for all choice set answered groups. Due to the different groups, there is even more change in attribute importance visible for the difference groups that answered difference amounts of choice sets. However, for group 3 there is a lower MAE compared to 4.2.1. Therefore, we can say that a higher WTR lead to less differences in attribute importance, so a negative moderation effect. For group 1 and 2 there is no or a positive moderating effect. So, we can accept H3a sincere there are changes in attribute importance’s for all groups. From table 8 we can infer the direct effect of WTR on attribute importance. We can see that the MAE is highest between 1-2 and lowest between 1-3. This means that a high WTR leads to lower differences in attribute importance and therefore we can reject H4a, so a higher WTR does not lead to more differences in attribute importance’s.

Table 7. Attribute importance’s image versus text

Less than 5,

n=102 More than 5, less than 16 n=78 More than 15, n=47 Attribute Image Text Difference Image Text Difference Image Text Difference Meat type 14,30% 13,53% 0,77% 17,35% 28,74% -11,39% 16,79% 22,83% -6,03% Cheese 11,18% 14,27% -3,08% 10,52% 5,13% 5,40% 11,88% 12,72% -0,84% Lettuce 30,58% 12,25% 18,33% 20,12% 10,75% 9,37% 22,61% 13,69% 8,91% Onions 0,15% 8,40% -8,25% 14,48% 9,89% 4,59% 6,16% 2,39% 3,77% Tomatoes 10,75% 8,09% 2,65% 13,81% 2,83% 10,98% 11,59% 9,81% 1,78% Sauce 6,01% 15,95% -9,94% 5,50% 22,00% -16,50% 8,69% 9,08% -0,40% Price 27,04% 27,51% -0,47% 18,22% 20,67% -2,45% 22,29% 29,48% -7,19% MAE 0,062 0,087 0,041

Table 8. Attribute importance’s image versus text

23 4.2.6 H3b and H4b: WTR on Image versus text on WTP

Again, as we have seen in 4.2.2 as well, there should be born in mind that there is a difference in price sensitivity and WTP. As can be observed in table 9 and 10 we have used two methods in order to estimate WTP estimates for the attribute levels. The first method is to estimate the possible moderating effect for WTR on WTP. The second method includes an interaction effect for attributes and sets answered and measures the direct effect of WTR on WTP. For both models, there is no difference in price sensitivity between the text condition and any other interaction. For method 1 we can see that there is a significant difference in WTP, in the image condition versus the text condition, for 1 patty of chicken, barbecue sauce (lower WTP), 1 patty of beef, lettuce and chili-sauce (higher WTP). We can interpret the WTP estimates like this; e.g. in the image condition respondents are willing to pay €1,26 extra for lettuce compared to the reference text condition, or €1,33 less for barbecue sauce. We expected that a higher WTR led to lower WTP and higher price sensitivity. We can clearly reject H4b, since neither the WTP is higher combined with a higher WTR, nor the price sensitivity changes.

Table 9. Marginal Willingness to Pay Estimates Image*Sets answered (Model 6)

Attributes Levels Text Image

Images*Sets answered

WTP p-value WTP p-value WTP p-value

Meat type Meat_Chicken_1x -1,03 *** 0,00002 -0,17 * 0,03344 -0,18 0,32367 Meat_Chicken_2x -0,33 0,13320 0,02 0,38164 0,01 0,49299 Meat_Beef_1x 0,08 0,70358 -0,55 0,14803 -0,55 0,76413 Meat_Beef_2x 1,28 *** 0,00010 0,70 ** 0,02000 0,72 ** 0,01000 Cheese Cheese_0x 0,00 1,00000 0,00 1,00000 0,00 1,00000 Cheese_1x 0,99 ** 0,00133 1,56 0,39797 1,57 0,95126 Cheese_2x 1,11 *** 0,00028 1,53 0,56097 1,53 0,77374 Lettuce No_Lettuce 0,00 1,00000 0,00 1,00000 0,00 1,00000 Lettuce 1,49 *** 0,00000 2,75 * 0,02229 2,77 0,39985 Onions No_Onions 0,00 1,00000 0,00 1,00000 0,00 1,00000 Onions 0,85 *** 0,00066 1,03 0,79727 1,03 0,68474 Tomatoes No_Tomatoes 0 1,00000 0 1,00000 0 1,00000 Tomatoes 0,60 * 0,01605 1,12 0,32249 1,14 0,26547 Sauce Sauce.Ketchup 0,10 0,63105 0,40 0,48579 0,38 0,19645 Sauce.Chili -1,01 *** 0,00002 0,21 ** 0,00299 0,21 0,92483 Sauce.Barbecue 0,87 *** 0,00002 -0,46 ** 0,00124 -0,46 0,96119 Sauce.Mayonaise 0,04 0,80000 -0,14 0,60000 -0,13 . 0,08000

Price Absolute estimates 0,38 *** 0,36 0,36

25

Table 10. Marginal Willingness to Pay Estimates Attributes and Attributes*Sets answered (Model 5)

Attributes Levels Attributes Attributes*Sets_answered

WTP p-value WTP p-value

Meat type Meat_Chicken_1x -0,38 0,1525 -0,39 0,1220

Meat_Chicken_2x -0,24 0,3617 -0,24 0,8970 Meat_Beef_1x 0,04 0,8652 0,03 . 0,0814 Meat_Beef_2x 0,57 * 0,0700 0,60 0,8000 Cheese Cheese_0x 0,00 * 0,0266 0,00 0,1770 Cheese_1x 1,09 0,2382 1,10 0,1166 Cheese_2x 2,18 0,0001 2,17 0,3000 Lettuce No_Lettuce 0,00 1,0000 0,00 1,0000 Lettuce 1,75 *** 0,0000 1,77 ** 0,0031 Onions No_Onions 0,00 1,0000 0,00 1,0000 Onions 1,13 *** 0,0002 1,12 0,4321 Tomatoes No_Tomatoes 0,00 1,0000 0,00 1,0000 Tomatoes 0,63 * 0,0364 0,65 * 0,0173 Sauce Sauce.Ketchup -0,23 0,4000 -0,22 1,0000 Sauce.Chili 0,59 * 0,0183 0,57 ** 0,0071 Sauce.Barbecue -0,71 ** 0,0091 -0,69 . 0,0758 Sauce.Mayonaise 0,34 0,1778 0,34 0,3143

Price Absolute estimates 0,35 *** 0,35

Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

4.2.7 H3c and H4c: WTR on Image versus text on Reliability

because there are different moderating effects for both conditions, a positive for text (accept H4c) and a negative for image. Regarding the direct effect of WTR on reliability we can see that the inverted U-shape holds, therefore we can accept H3c.

Table 11. Reliability among different groups

Chi-square

Group 1: <8 Out of: Percentage P-value

Total 5 7 71,43%

Image 1 2 50,00%

Text 4 5 80,00% 0,427

Group 2: >7, <16 Out of: Percentage:

Total 29 36 80,56%

Image 15 19 78,95%

Text 14 17 82,35% 0,797

Group 3: >15 Out of: Percentage:

Total 36 50 72,00%

Image 19 32 59,38%

Text 17 18 94,44% 0,008 **

Total all consistent Out of: Percentage:

Total 70 93 75,27% Image 35 53 66,04% Text 35 40 87,50% 0,018 * Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% 90.00% 100.00%

Group 1 Group 2 Group 3

P er ce nt age c o rr ec t

Number of Choice Sets Answered

Reliability

Total Image Text

27

5. Discussion

This article reports on the results from a series of experiments on the impact of text versus image on accuracy, WTP, reliability and WTR. Also, it shines light on whether the number of choice sets answered influences the relationship between images versus text on the accuracy, WTP and reliability. At last we researched whether there was a direct effect of WTR on accuracy, WTP and reliability. We estimated most of the models by a multinomial logit regression. The task complexity was kept constant, but the design varied between the image condition and the text condition. By using voluntary choice sets this study provides individual-level estimates of the number of choice sets a consumer is willing to complete.

We found that there are differences in attribute importance’s between the text and image condition, this is in line with Anderson (1987); Louviere et al. (1987); Smead et al. (1981) and Vriens et al. (1998). We as well found that judgements are easier made in the image condition, leading to different estimates. There is no difference in price sensitivity, but WTP is higher in the image condition, this is in line with Umberger and Mueller (2010), there is underestimation of the price attribute in the text condition. Also, respondents have a higher WTR in the image condition, mainly due lower depletion of resources, however, the text condition led to more reliable answers, this is not in line with Dellaert, Donkers and Soest (2012); Haaijer, Kamakura and Wedel (2000) and Otter, Allenby and van Zandt (2008), it might well be that in the text condition, that takes more effort, respondents put more effort in considering the choices, leading to more reliable answers.

reliability, but a negative moderating effect for the image condition. The positive moderating effect of WTR on reliability in the text condition is contradicting current research. Regarding accuracy, price sensitivity and WTP, WTR does not have an influence.

5.1 Managerial Implications

Managers and marketing researchers can learn from this results that the optimal number of choice sets of between 8 and 15 for an image based choice experiment, whereas for the text design reliability is highest for over 15 choice sets. Regarding pricing is also of importance to consider the differences between image and text choice set design, there is on average a higher WTP in the image condition. Also, if you would like to have a high response rate one should invest in choice sets that contain images, this will lead to the lowest fatigue. Also, for the attribute importance’s there are differences between images and text as well as for different number of choice sets answered, so that has to be taken into account as well. At last, this paper shows that there are no differences for price for different number of choice sets answered.

5.2 Limitations and Future Research

29

Reference List:

Alberini, A. (1995). Efficiency vs Bias of Willingness-to-Pay Estimates: Bivariate and Interval-Data Models. Journal of Environmental Economics and Management.

Alphonce, R., & Alfnes, F. (2012). Consumer willingness to pay for food safety in Tanzania: An incentive-aligned conjoint analysis. International Journal of Consumer Studies, 36(4), 394– 400.

Anderson, J. C. The effect of type of representation on judgements of new product acceptance.

Industrial Marketing and Purchasing 2:29–46 (1987).

Bech, M., Kjaer, T., & Jorgen, L. (2011). Does the Number of Choice Sets Matter? Results From a Web Survey Applying a Discrete Choice Experiment. Health Economics, 20, 273–286. Bradley, M., & Daly, A. (1994). Use of the logit scaling approach to test for rank-order and fatigue

effects in stated preference data. Transportation, 21(2), 167–184.

Carlsson, F., & Martinsson, P. (2008). How much is too much? An investigation of the effect of the number of choice sets, context dependence and the choice of bid vectors in choice

experiments. Environmental and Resource Economics, 40(2), 165–176.

Carlsson, F., Mørkbak, M. R., & Olsen, S. B. (2012). The first time is the hardest: A test of ordering effects in choice experiments. Journal of Choice Modelling, 5(2), 19–37.

Caussade, S., Ortúzar, J. de D., Rizzi, L. I., & Hensher, D. A. (2005). Assessing the influence of design dimensions on stated choice experiment estimates. Transportation Research Part B:

Methodological, 39(7), 621–640.

Chung, C., Boyer, T., & Han, S. (2010). How Many Choice Sets and Alternatives are Optimal? Consistency in Choice Experiments. Agribusiness, 22(2), 175–199.

Czajkowski, M., Giergiczny, M., & Greene, W. H. (2014). Learning and Fatigue Effects Revisited: Investigating the Effects of Accounting for Unobservable Preference and Scale

Heterogeneity. Land Economics, 90, 324–351.

Day, B., Bateman, I. J., Carson, R. T., Dupont, D., Louviere, J. J., Morimoto, S., Wang, P. (2012). Ordering effects and choice set awareness in repeat-response stated preference studies.

Journal of Environmental Economics and Management, 63(1), 73–91.

Dellaert, B. G., Donkers, B., & Soest, A. Van. (2012). Complexity Effects in Choice Experiment– Based Models. Journal of Marketing Research, 49(3), 424–434.

DeShazo, J. R. R., & Fermo, G. (2002). Designing choice sets for stated preference methods: The effects of complexity on choice consistency. Journal of Environmental Economics and

Management, 44(1), 123–143.

Ding, M., Grewal, R., & Liechty, J. (2005). Incentive-Aligned Conjoint Analysis. Journal of

Marketing Research, 42(1), 67–82.

Domzal, T. J. and Unger, L. S. Judgments of verbal versus pictorial presentations of a product with functional and aesthetic features. Advances in Consumer Research 12:268–272 (1985). Eggers, F. (2015). The Number of Choice Sets in Conjoint Choice Experiments: An Analysis of

Willingness-to-respond, 1–7.

Eggers, F., & Sattler, H. (2011). CBC HIT - CBC HILCA Preference Measurement with Conjoint Analysis Overview of State-of-the-Art Approaches and Recent Developments, 3(1), 36–47. Erdem, T., Swait, J., & Louviere, J. (2002). The impact of brand credibility on consumer price

sensitivity, 19(1), 1–19.

Grainger, J., Kiyonaga, K. & Holcomb, P. J. (2006). The time course of orthographic and phonological code activation. Psychological Science, 17, 1021–1026.

Haaijer, R., Kamakura, W., & Wedel, M. (2000). Response Latencies in the Analysis of Conjoint Choice Experiments. Journal of Marketing Research, 37(3), 376–382.

Hauser, J. R., Eggers, F., & Selove, M. (2016). The Strategic Implications of Precision in Conjoint Analysis by The Strategic Implications of Precision in Conjoint Analysis Abstract.

Hensher, D. A. (2004). Identifying the Influence of Stated Choice Design Dimensionality on Willingness to Pay for Travel Time Savings Institute of the Australian Key Centre in Transport Management the University of Sydney, 38(April 2004), 425–446.

Hensher, D. A. (2006). Revealing differences in willingness to pay due to the dimensionality of stated choice designs: An initial assessment. Environmental and Resource Economics, 34(1), 7–44. Holbrook, M. B., & Moore, W. L. (1981). Feature Interactions in Consumer Judgments of Verbal

versus Pictorial Presentations. Journal of Consumer Research, 8(1), 103–113.

Jensen, E. (2008). Brain-based learning: The new paradigm of teaching (2nd ed.). Thousand Oaks, CA.: Corwin Press.

Johnson, R. M., & Orme, B. K. (2002). How Many Questions Should You Ask in Choice-Based Conjoint Studies? Beaver, 98382(360), 1–24.

Levie, W. H. & Lentz, R. (1982). Effects of text illustrations: a review of research. Educational

Communication and Technology, 30, 195–232.

Louviere, J. J., Schroeder, H., Louviere, C. H., & Woodworth, G. C. (1987). Do the parameters of choice models depend on differences in stimulus presentation: visual versus verbal presentation. Advances in Consumer Research, 14(1), 79–82.

Manski, C. F. (1977). The Structure of Random Utility Models. Econometrica, 8(3), 229–254.

McFadden, D. (1974), ‘Conditional logit analysis of qualitative choice behavior’, in P. Zarembka, ed., Frontiers in Econometrics, Academic Press, New York, pp. 105– 142.

Orme, B., 2010a. Getting Started with Conjoint Analysis: Strategies for Product Design and Product Research, Second Edition. Research Publisher Madison.

Otter, T., Allenby, G. M., & van Zandt, T. (2008). An Integrated Model of Discrete Choice and Response Time. Journal of Marketing Research, 45(5), 593–607.

Pieters, R., & Warlop, L. (1999). Visual attention during brand choice: The impact of time pressure and task motivation. International Journal of Research in Marketing, 16(1), 1–16.

Pocheptsova, A., Amir, O., Dhar, R., & Baumeister, R. F. (2009). Deciding Without Resources: Psychological Depletion and Choice in Context. Journal of Marketing Research, Vol. 46 (No. 3), 344–355.

Potter, M. C., Wyble, B., Hagmann, C. E. & McCourt, E. S. (2014) Detecting meaning in RSVP at 13 ms per picture. Attention, Perception, and Performance 76(2):270–79.

Ryan, M., & Farrar, S. (2000). Using conjoint analysis to elicit preferences for health care. BMJ

(Clinical Research Ed.), 320(7248), 1530–1533.

Smead, R. J., Wilcox, J. B., & Wilkes, R. E. (1981). How Valid are Product Descriptions and Protocols in Choice Experiments? Journal of Consumer Research, 8(1), 37–42.

Townsend, C., & Kahn, B. E. (2014). The “Visual Preference Heuristic”: The Influence of Visual versus Verbal Depiction on Assortment Processing, Perceived Variety, and Choice Overload.

Journal of Consumer Research, 40(5), 993–1015.

Umberger, W. J., & Mueller, S. C. (2010). Is Presentation Everything? Using Visual Presentation of Attributes in Discrete Choice Experiments to Measure the Relative Importance of Intrinsic and Extrinsic Beef Attributes. Agricultural & Applied Economics Association 2010

AAEA, CAES, & WAEA Joint Annual Meeting, 1–29.

Vriens, M., Loosschilder, G. H., Rosbergen, E., & Wittink, D. R. (1998). Verbal versus realistic pictorial representations in conjoint analysis with design attributes. Journal of Product

31

Appendix

Table 12. Chi-Square Test for Reliability

Text Image Total

Reliable 35 35 70 Unreliable 5 18 23 Total 40 53 93 Expected value 30,11 39,89 70 9,89 13,11 23 Total 40 53 93 p-value 0,017546769 * Significant codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Table 13. T-test sets answered

Welch Two Sample t-test data: dat$Sets_answered by dat$Images t = -1,9124 df = 234,07 p-value = 0,05705

alternative hypothesis: true difference in means is not equal to 0

95 percent confidence interval:

-6,180514 0,091964

sample estimates: mean in group 0 mean in group 1

9,475410 12,519690

Table 14. T-test log transformed sets answered Welch Two Sample t-test

data: log(dat$Sets_answered) by dat$Images t = -1,9376 df = 243,83 p-value = 0,05382 alternative hypothesis: true difference in means is not equal to 0

95 percent confidence interval:

-0,394444 0,003242

sample estimates: mean in group 0 mean in group 1

Table 15. Output of the Cox Proportional Hazard Model

coef exp(coef) se(coef) z Pr(>|z|)

Lecture -0,549938 0,576986 0,228468 -2.407 0,01608 * Images -0,609137 0,54382 0,206149 -2.955 0,00313 ** Gender 0,270327 1.310.393 0,215328 1.255 0,20933 Age -0,008358 0,991677 0,012713 -0,657 0,5109 X74a -0,058749 0,942944 0,097543 -0,602 0,54698 X74b -0,110022 0,895815 0,089471 -1.230 0,21881 X74c 0,03363 1.034.202 0,092411 0,364 0,71592 X74d -0,074667 0,928052 0,082836 -0,901 0,36738 X74e -0,033871 0,966696 0,105228 -0,322 0,74754 X75a -0,076877 0,926004 0,108144 -0,711 0,47716 X75b -0,0847 0,918788 0,095673 -0,885 0,37599 X75c -0,019959 0,980239 0,107172 -0,186 0,85226 X75d 0,231683 1.260.719 0,101273 2.288 0,02215 * X76a 0,038205 1.038.945 0,083263 0,459 0,64634 X76b -0,161424 0,850931 0,093745 -1.722 0,08508 . X76c -0,051042 0,950239 0,074059 -0,689 0,49069 X76d 0,129366 1.138.107 0,083135 1.556 0,11968 X76f -0,023647 0,976631 0,110327 -0,214 0,83029 Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Concordance= 0.623 (se = 0.033 )

Rsquare= 0.091 (max possible= 0.985 ) Likelihood ratio test= 23.8

on 18 df p=0.1617

Wald test = 23.57

on 18 df p=0.1695

Score (logrank) test =

23.83 on 18 df p=0.1606

33

Coefficients Estimate Std. Error t-value Pr(>|t|) Meat_Chicken_1x -0,2471682 0,0545419 -45.317 5,85E-03 *** Meat_Chicken_2x -0,0989727 0,0525011 -18.852 0,05941 . Meat_Beef_1x -0,1205274 0,0537109 -22.440 0,02483 * Cheese.layers 0,7350880 0,1293878 56.813 1,34E-05 *** I(Cheese.layers^2) -0,2521064 0,0596073 -42.295 2,34E-02 *** Lettuce 0,8722699 0,0626372 139.257 <2,2e-16 *** Onions 0,3138161 0,0601297 52.190 1,80E-04 *** Tomatoes 0,4267321 0,060534 70.495 1,80E-09 *** Sauce.Ketchup 0,0078223 0,0512448 0,1526 0,87868 Sauce.Chili -0,1123262 0,0524916 -21.399 0,03236 * Sauce.Barbecue 0,0479667 0,0520532 0,9215 0,35679 Price.6 0,5277504 0,0475354 111.023 <2,2e-16 *** Price.7 0,1105946 0,0507752 21.781 0,0294 * Price.8 -0,1016327 0,0534733 -19.006 0,05735 . None_option 2,0140751 0,0954418 211.027 <2,2e-16 *** Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Log-Likelihood: -2749.1 Table 18. Raw Output Model (3)

I(Images*Lettuce) 0,54239 0,12849 4,22130 0,02429 *** I(Images*Onions) -0,01911 0,12376 -0,15440 0,87726 I(Images*Tomatoes) 0,32070 0,12421 2,58190 0,00983 ** I(Images*Sauce.Ketchup) -0,03961 0,10554 -0,37530 0,70740 I(Images*Sauce.Chili) 0,44889 0,11166 4,02030 0,05811 *** I(Images*Sauce.Barbecue) -0,50562 0,10599 -4,77050 0,00184 *** I(Images*Price.6) -0,04834 0,09786 -0,49400 0,62131 I(Images*Price.7) -0,12364 0,10403 -1,18850 0,23464 I(Images*Price.8) 0,06039 0,11083 0,54490 0,58585 I(Images*None_option) 0,31499 0,19324 1,63010 0,10308 Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 Log-Likelihood: -2711.3

Table 19. Raw Output Model (5)

35

Coefficients Estimate Std.Error t-value Pr(>|t|)

Table 21. Attribute importance's 2 methods Model (3) Hierarchical Bayes Difference