Conjoint analysis with stated choice probabilities:
Studying choice uncertainty change in time for innovative
products in the smart watches market
Yordanka Gyurdieva
Student number: 3011267
University of Groningen
Faculty of Economics and Business
Master Thesis, Msc Marketing Intelligence
6/20/2016
1 | P a g e
CONTENTS
1. INTRODUCTION ... 3
2. RESEARCH FRAMEWORK ... 4
2.1 CONJOINT ANALYSIS ... 4
2.2 UNCERTAINTY IN INNOVATIVE PRODUCTS ... 7
3. METHODOLOGY ... 8 3.1 CONCEPTUAL MODEL ... 8 3.2 HYPOTHESIS FORMULATION ... 10 3.3 ESTIMATION PROCEDURE ... 11 4. DATA ... 12 4.1 STUDY DESIGN ... 12 4.1.1 INDEPENDENT VARIABLES ... 12 4.1.2 DEPENDENT VARIABLE ... 13 4.1.3 MODERATOR ... 14 4.1.4 COVARIATES ... 15 4.1.5 CHOICE DESIGN ... 15 5. RESULTS ... 16 5.1 DESCRIPTIVE STATISTICS ... 17
5.2 LINEAR OR NOMINAL ATTRIBUTES ... 19
5.3 MODEL FIT: LIKELIHOOD RATIO TEST ... 21
5.4 HYPOTHESIS EXAMINATION ... 22 5.4.1 HYPOTHESIS 1 ... 22 5.4.2 HYPOTHESIS 2 ... 23 5.4.3 HYPOTHESIS 3 ... 24 6. DISCUSSION ... 26 7. LIMITATIONS ... 27 8. CONCLUSION ... 28 REFERENCES ... 29 APPENDIX ... 32
TABLE 1 – ATTRIBUTES AND LEVELS ... 32
2 | P a g e
TABLE 3 – SURVEY 1st PART POPULATION MOMENTS ... 34
TABLE 4 – SURVEY 2nd PART POPULATION MOMENTS ... 35
TABLE 5 – STANDARD DEVIATION OF POPULATION MOMENTS ... 35
PLOT 1- SURVEY 1st PART –INDIVIDUAL ATTRIBUTES IMPORTANCE ... 37
PLOT 1- SURVEY 2nd PART – INDIVIDUAL ATTRIBUTES IMPORTANCE ... 38
PICTURE 1 – SURVEY SCREENSHOTS ... 39
PHASE ONE ... 39
3 | P a g e
1. INTRODUCTION
The present study aims at investigating the uncertainty in consumer’s preferences for innovative products by using choice probabilities as an input in conjoint analysis. In particular, the market for smart watches was examined since it can be considered a representation of a product which makes consumers inexact in their preferences when being considered an accessory up until now.
4 | P a g e
2. RESEARCH FRAMEWORK
In the following section the literature background of the particular topic will be introduced together with an explanation of the reason for choosing the market for innovative products.
2.1 CONJOINT ANALYSIS
Conjoint analysis is a tool, vastly applied in Marketing research, which foundations date back to 1920s but it is agreed among scholars that the paper of Luce and Tukey, published in 1964, concerning the theory of conjoint measurement, set the start of the analysis’ applications (Rao & Vithala, 2014; Green & Srinivasan, 1978). It has been improved and used to analyze trade-offs made by the consumers ever since. Understanding your customer is a crucial pinpoint in every industry or business, and conjoint analysis may provide many answers which could make the whole process easier, if used properly. It is a tool for analyzing consumers’ preferences and developing an optimal product, segmenting the market, computing the maximum price consumers are willing to pay and providing answers to many other managerial questions. But probably the most important feature of the analysis is that it could be used as a basis and a tool for predicting consumer behavior. The core idea behind it is that consumers evaluate the product according to the utility it brings to them and therefore, the analysis assumes that this utility can be decomposed into utilities of each component or attribute of the product and their interactions, if present (Rao & Vithala, 2014). Moreover, quite frequently consumer’s behavior involves processes which are outside of their conscious awareness. Furthermore, sometimes consumers are not aware of the features that trigger certain kind of behavior (Chartrand, 2005). In other words, simply asking the consumers which is their most preferred product feature and which attribute makes them want to buy it is not going to produce accurate results every time, which is where conjoint analysis creates value when understanding consumer preferences is concerned.
5 | P a g e
making process. In the past, ratings- or rankings- based data were the top choice and most frequently used, which are collected by asking the respondents to rank the presented alternatives in order of most to least preferred product. Later, the choice-based conjoint analysis became popular as an alternative for measuring preferences. (Vriens et al. 1998) It was one of the most significant developments in conjoint analysis in the 80’s (Green et al., 2001). For this type of conjoint analysis, the data is collected by asking the respondents to mark a single preferred choice. It is defined by Louviere (1988), as “responses that identify one and only one of a set of alternatives as the “highest”, “best”, etc.” As compared to ratings- and rankings- based conjoint analysis, it was considered as a more realistic representation of reality because of its simplicity and similarity to a real purchasing decision. (Desarbo et al., 1995) In real life people are not expected to rank or order alternatives before choosing one to purchase, instead they just pick one. In cases when data on real purchasing choices is not available, the collected data on stated choices is used for the estimation of utility models.
However, as it is common in similar analyses, the choice-based conjoint analysis has its drawbacks and critics. Having a simplistic input data (single-choice) inevitably leads to assumptions one has to make about the psychological and physical processes influencing or leading to that decision. In other words, a consumer answering to a “yes / no” questions is like a statistician making a best point prediction of a future random event (Manski, 2004). Therefore, this is the most probable answer currently but there are other possible end results. Jsuter (1966) first reasoned that asking the respondents for purchase probabilities will be more informative than buying intentions. More to the point, he proposed asking questions that associate verbal expressions of likelihood with numerical probabilities. Further data collection and analysis led to proving the superiority of eliciting choice probabilities over “yes / no” answers. At that time, however, econometricians were focused on single choice data and this discovery received attention around a quarter of century later. (Manski,2004)
6 | P a g e
give an inaccurate prediction since, in a real choice setting, customers will acquire and be exposed to much more information about the product of interest. The solution is a different type of data, collected by asking the respondents to allocate probabilities of purchasing behavior to the different options presented. In other words, as put by Ajzen and Fishbein (1980), the “behavioral intention” of a person is the subjective probability that he or she allocates to the behavior of interest actually occurring. This approach allows for respondents’ uncertainty and elicits preferences mixed with the expectation of acquiring more information (Manski, 1999). Strictly speaking, the decision-maker will use available information and make an uncertain decision based on the subjective expectations which will maximize their expected utility. In choice-based conjoint analysis the possibility of objective probability being different from subjective expectations, which is often the case, is overlooked. (Delevande, 2008).
7 | P a g e
2.2 UNCERTAINTY IN INNOVATIVE PRODUCTS
The common understanding of the word “new” is usually something unfamiliar, uncertain and therefore risky. When it comes to new products on the market the ambiguity comes from the limitation of consumer’s knowledge on the usage, purpose and value created by individual characteristics and their combination of the product itself.
Litter and Melanthiou (2006), defines uncertainty, regarding the adoption of new product or service, as doubt expressed by the consumers, on many aspects such as what they are seeking in the new product, the sources and type of information they are using when making the decision, the credibility of the brand, the criteria needed for product evaluation and the range of alternative offers they should consider. In other words, there are so many unknowns that it is hard for the consumer to make decision on buying a single product. One of the reasons behind is that since the product is new on the market, there is not enough information concerning its usage to reduce this uncertainty. Feedback from early adopters will increase the trust of the potential consumers in the product characteristics and usability, also enrich their knowledge of the product and therefore make them more confident in answering all of the questions mentioned previously.
Moreover, the lack of information on usage is related to doubt in the product liking or preferences and decreases willingness to pay for the product (Kim & Krishnan, 2015). The reason behind it is the low level of consumer confidence in their knowledge of the product not only because they are not able to choose one to buy, but because not being informed enough means that one is not able to effectively evaluate the true merits of the product, or in other words what value it brings to them and therefore what is, in their opinion, a reasonable pricing of that product / service.
8 | P a g e
for them. Finally, since the smartwatch is relatively expensive innovation on the market, people will not be willing to buy it until they are confident in their knowledge of the product which means they are very likely to enrich their understanding with additional information from an outside source.
Having this in mind, the market for smart watches is full of ambiguity and therefore is suitable for studying the change in consumer’s uncertainty with time and acquisition of additional information by eliciting choice probabilities. Moreover, little literature is devoted to examining the analysis of uncertainty by eliciting choice probabilities and using it in conjoint analysis.
3. METHODOLOGY
In this section the conceptual model of the study is presented together with stating the hypothesis which are about to be tested and the estimation procedure used for the purpose.
3.1 CONCEPTUAL MODEL
In order to investigate the effect of time and new information on consumers’ choice, first a model should be specified. In order for a relationship to be defined, independent variables which have an effect on some dependent variable have to be present. In the case of conjoint analysis, the uncertain parameter, that the model is trying to estimate and predict, is the preference and choice of the consumer. The probabilities allocated to each product by the respondents, as representative of choice, are the dependent variable. The independent variables are the ones which are expected to have an effect on the dependent variable. Therefore the attributes of the product – the smart watch, are expected to influence preference of the consumer. The underlying assumption is that based on the attributes of each alternative, the respondents allocate subjective probabilities to each alternative, so that the choice maximizes their expected future utility (Delevande, 2008).
9 | P a g e
most influence over determining consumer’s preferences are: brand, price, type of the band and extras. The attributes’ importance is not studied separately, instead the combination of different levels among the attributes, as presented to the respondent, are assessed. Therefore, measuring their relative importance is the final result.
Moreover, some control variables are included, as covaiates, for better understanding of individual characteristics of the respondents in the sample used. They will be stated and explained in the study design.
As mentioned previously, time is considered to have a major impact on consumer’s preferences because of the additional information they can acquire before the actual purchase. Due to the limitation of time and resources for the data collection of this study, a proxy will be used for time, which will, again, be explained and motivated in the study design part. It will be included as a moderator of the relationship between the dependent variable, represented by the elicited choice probabilities, and the independent variables – the different combinations of attributes. Given the collected data, the presence of its effect will be best observed in the changes of the allocated probabilities.
10 | P a g e
Figure1 – Conceptual model
3.2 HYPOTHESIS FORMULATION
Three hypotheses are going to be specified concerning the change in preferences with time as measured in the probabilistic choice model. Firstly, uncertainty about a single most preferred product is expected to be present in the form of respondents’ answers. They are expected to allocate different percentage probabilities, avoiding extremes of 0% and 100%, due to uncertainty and limited information. As suggested by Manski (1999), the information provided in the survey will be limited, the respondents will know it and will expect to enrich their knowledge about the product in the future, before the actual purchase, which will cause doubt in the provided options, especially in the first phase of the survey. Additionally, this will be evidence of the superiority of using choice probabilities over stating a single choice, in innovative products scenarios.
11 | P a g e
is related to increasing indecisiveness or heterogeneity in respondent’s preferences because it creates additional aspect for differentiation of preferences and difference in interpreting product features and merits (Malhotra, 1982). However, the additional variable in this case, battery life, is not expected to create additional uncertainty because it is not ambiguous in the value it creates or unfamiliar to the respondents.
Last but not least, the error variance of the model is expected to increase with time due to the increase in heterogeneity of preferences which results from confidence of the respondents in their choice. Apart from presented with timeline and exposed to additional product information, the literature suggests that the repeated exposure of the respondents to the same attributes with different levels increases knowledge of the benefits available from each and the familiarity with the choice task. As a result the respondents become more determined in their choice. (Huber et al., 2002) Therefore, the increase in understanding and knowledge of the product is expected to help the respondents better define their preferences which will lead to greater observed heterogeneity. Therefore, the specified, aggregated model will increase its error variance because it will be worse representation of the data.
3.3 ESTIMATION PROCEDURE
Many researchers agree that long and information-flooded questioners result in respondents not presenting their real preferences and therefore, a wrong model results. However, a small number of observations per respondent will also lead to less realistic estimates. The solution for both of those problems is called a Hierarchical Bayes model.(Lenk et al.,2016) Individual coefficients of the respondents were achieved via a Monte Carlo Markov Chain method with a 5000 iterations and each 5 kept. Next the means and standard deviations for all attributes per respondents were taken as the individual parameters. Additionally, those coefficients’ mean and standard deviation were taken to specify the coefficients of an aggregate model. The heterogeneity in preferences, as well as the aggregate estimates, and their change in time were needed for the purpose of examining the earlier stated hypothesis.
12 | P a g e
Yi = Xi * βi +ɛi for i=1…,n (1) βi= Θ*Zi + δi for i=1…,n (2)
In equation (1) Y is a vector of the probabilistic responses of subject i. X is the effect-coded design matrix which describes the i-th survey version presented to the i-th respondent. The estimated regression coefficients are stored in the vector β, for each individual. Equation (2) is a description of the heterogeneity of individual preferences by accounting for the covariates ( in this case demographic characteristics and knowledge of the product presented). Zi are the mean-centered covariates and Θ is a matrix of regression coefficients. ɛiand δi are the error terms of both equations with extreme value of type one and a multivariate normal distributions, respectively.
4. DATA
In this section, the data needed for the study will be introduced with its structure and purpose of the different variables. Additionally, the design of the survey will be explained in detail.
4.1 STUDY DESIGN
4.1.1 INDEPENDENT VARIABLES
13 | P a g e
in that market – Apple and Samsung. However, in order to keep the number of levels the same across all attributes, the brand Asus will also be included, representative of the relatively cheaper smartwatches. The third variable included is the band type, which is a very important characteristic of the watch because wristwatches have long been considered an accessory and therefore the way they look is important. Many seems to think that this is one of the reasons smartwatches are not yet viral in the electronics market (Pogue, 2014). There seems to be only one conjoint analysis study done on smartwatches, as of my knowledge. It was done by students in India in 2012 as part of a Marketing Research course. They found out that the color of the band is the second most important attribute of the smartwatch after the price. This can be considered as support for the previous claim. Therefore, the type of band is included in the analysis with three levels – leather, metal and rubber. The fourth feature included is extras with, again, three levels – GPS, Wi-Fi and both GPS and Wi-Fi. Last but not least, fitness extras are included. The reason behind the inclusion of this attribute is that the main advantage of the smartwatch over the smartphone is considered that it is constantly in contact with the person’s body and therefore, can more precisely track body motion (Pogue, 2014). According to the customer reviews these are also one of the top characteristics in terms of importance, especially among physically active consumers. The three levels are heart-rate monitor with compass, heart-rate monitor with accelerometer and accelerometer, pedometer and a compass. In the second phase of the survey, the additionally added attribute is battery life with values of 8 hours, 10 hours and 15 hours. This attribute was chosen so that it is familiar to the respondent and they understand its value and therefore does not create additional uncertainty. A summary of the attributes and their levels is presented in table 1 in the Appendix.
4.1.2 DEPENDENT VARIABLE
14 | P a g e
choice as a probability of buying the three different options. The four probabilities will have a sum of 100%. The elicited choice-probabilities will be the dependent variable.
4.1.3 MODERATOR
15 | P a g e
4.1.4 COVARIATES
At the beginning of the survey respondents will be asked some questions for collecting the control variables needed. Age, gender and education level may all affect preferences. More importantly, monthly budget is included as a question so that the price of the smartwatch can be taken as percentage of it and possibly make the end results of the study generalizable to buying behavior of such innovative durables that increase involvement because of relatively high price. To make sure the respondents are representative of the average consumers it is important for them not to be resistant to adopt new technology and understand the product presented. Thus, their innovation adoption willingness and knowledge of the smart watch as a product are measured on 5-point Likert scale.
4.1.5 CHOICE DESIGN
The survey was created using the Sawtooth Software. A fractional factorial design was used due to the high number of possible attribute combinations. The design is orthogonal and balanced with minimum overlap, meaning that the attribute levels and attribute level pairs appear equal number of times in order not to bias the results. Additionally, dominating alternatives and inconsistent alternatives were not included.
When it comes to the choice sets, three alternatives are presented to the respondent. Since the survey is divided into two parts, each part consists of 7 choice sets to be evaluated. The reason is that in order for the respondents to be effectively evaluating the alternatives, without fatigue, they should be provided with additional stimuli if the choice sets exceed 15. (Eggers & Sattler, 2011) Additionally, in the software guidelines it is written that a number between 6 and 10 choice sets are needed for the estimation of a model.
The survey starts with 5 questions to collect data on the control variables mentioned previously. Next, in order to be certain in the respondents’ understanding of how the choice probabilities work an explanation is added before the beginning of the survey saying the following:
16 | P a g e
chance to buy the first watch, 20% chance to buy the second one and 30% chance to buy the last one. You can also allocate probabilities such as 100%, 0%, 0%.”
Following 7 choice sets are presented for evaluation. The choice sets include a picture and the different characteristics of each product, represented by the attributes and their levels specified earlier. Before the second phase an explanation of what is going to happen next is added so that an attention on the additional information provided is ensured. The explanation states the following: “Now imagine you are in the store right now and you remember seeing the ad presented below (please take a minute to watch it). However, you are curious about the battery life of each watch and you use your mobile phone to find the information online. The battery life will be presented for each watch in the next few questions. Please allocate probabilities for each product again.” In both descriptions (before the first and the second phases) the point in the timeline was bolded to that it captures attention and is stressed upon. Next, a one-minute ad of a smartwatch is played without the last few seconds which show the brand name in order not to bias the results. After that additional 7 choice sets are presented for evaluation with the additional attribute – battery life, mentioned in the description.
Additionally, in order for the survey to better represent reality, a no-choice option was added as a fourth alternative in every choice set to which the respondents had to allocate probability as well. Using a “no-choice” option increases the design efficiency and better mimics the choice process. Screen shots of a choice sets are presented for the first and second phases of the survey in Picture 1 of the Appendix.
5. RESULTS
17 | P a g e
5.1 DESCRIPTIVE STATISTICS
Out of 156 respondents, 118 have successfully completed the first part of the survey and 114, the second one which results in approximately 76% completion rate. The average time for completion was 9 minutes with a standard deviation of 4 minutes. Every respondent was shown 8 choice sets, followed by a video of a smartwatch and an additional 8 choice sets with one more attribute. Each choice sets constituted of 4 alternatives, one of which was a no-choice option.
The sample is almost equally divided in terms of gender with 56% male and 44% female. The respondents’ age ranged between 20 and 50 with an average of 24. In terms of education level, all of the participants had a university degree with 46% having/working on their Bachelor’s degree and 54% having/working on their Master’s degree.
18 | P a g e
Additionally, their innovativeness was examined by asking to choose a level ( on a scale of 1-not likely to 5-very likely) their likeliness of buying innovative products. As presented in Graph 2 below, much greater part of the respondents are willing to adopt new technology showing that they are a good sample to test the smartwatches preferences on.
Graph2
Last but not least, the respondents were asked how familiar they are with the product, again on a scale from 1 to 5. As presented in Graph 3 below, greater part of them knew about the existence of the product and many were very well informed about it. Therefore, the results will not be influenced by lack of knowledge which, in case of being present, may cause the respondents not providing real preference but randomly answering.
19 | P a g e
Graph 3
5.2 LINEAR OR NOMINAL ATTRIBUTES
In a conjoint study the different attributes may take on a different form meaning that sometimes there might be a specific linear relationship between the different attribute levels and their utilities or each such pair may be unrelated to the others. In other words, it is not compulsory for all the attributes to be included as either part-worth or linear but a combination can be used depending on logic and model fit. In the case of the smartwatches, most the included attributes and their levels seems to have no particular relationship since they are qualitative. However, when it comes to price the rule of thumb says higher price mean lower demand. Therefore, price may be better included as a linear variable.
In order to examine if price should be included as a linear attribute both a graph and model likelihood will be considered. In Graph 4 below, the aggregate price coefficients are plotted against their respective level. As expected they appear to have a linear relationship.
0% 5% 10% 15% 20% 25% 30% 35% 1 2 3 4 5
20 | P a g e
Graph 4
This linear relationship is tested by estimating both part of the survey (representative of the two points in time) with all attributes as part-worth once and a second time by including price as a linear variable. In the case of Hierarchical Bayes model the Newton-Raftery approximation of the log marginal density had to be used in order to reach a single loglikelihood measure. The first 599 draws were used as burn-in and log likelihood of the last 400 we used as an input to the function. In the table below, the numbers for the two parts of the survey with all attribute as part-worths and all as part-worths but the price, are presented. In both cases, for the first and second part of the survey, adding price as a linear variable results in a lower marginal density of the log likehoods.
. Log marginal density
1st part - all part-worths -1225.033
1st part - price linear -1216.769
2nd part - all part-worths -1099.33
2nd part - price linear -1079.606
The difference in the loglikelihood presented is very small and it might be due to the change in the number of parameters. For that reason, the adjusted R2 is computed and examined as well. The computations were done using the formula for the McFadden adjusted R2:
21 | P a g e
, where LL(0) is the log likelihood of the NULL model.
The two tables below presents the results for the two parts of the survey. The adjuster R2 again indicates that a model with the price included as a linear variable is a better fit to the data. Therefore, price will be included as a linear variable.
1st part survey npar Adj R sq LL(0) -1308.66188 LL(b)price-pw -1225.033 11 0.05549858 LL(b)price-linear -1216.769 10 0.06257757 2nd part survey npar Adj R sq LL(0) -1264.30046 LL(b)price-pw -1099.33 13 0.12020122 LL(b)price-linear -1079.606 12 0.13659289 5.3 MODEL FIT: LIKELIHOOD RATIO TEST
22 | P a g e
First part survey .
LL(0) -1308.6619 LL(b) -1216.769 npar (b) 10 respondents (n) 118 choice sets ( c) 8 alternatives (m) 4 Chisq 183.785754 p-value <0.0001 5.4 HYPOTHESIS EXAMINATION 5.4.1 HYPOTHESIS 1
In line with the expectations, uncertainty in the choice of product is present in both phases of the survey, expressed by different probabilities allocated to the available options. Histograms, for both parts of the survey, of frequencies of the different values of the selection variable are presented in Table 2 in the Appendix. Different values are used, as can be seen in histograms, with both a single answer present (in the form of a 100%) and an uncertainty, expressed by allocation of probability choice. In the first part of the survey choices of 50%, 30% and 20% are allocated by the respondents more often than a 100%. A specific choice, stated by the allocation of a 100% to a single option, is more frequently present in the second part of the survey which was expected since at that point in time the respondents had more information about the product which made them more certain of their choice. However, even when that is the case different probabilities were allocated in many cases, again, indicating uncertainty. 10%,20% and 50% are choices often met in the data.
All of the abovementioned is an indication of uncertainty which must be accounted for in the conjoint studies to make them more precise, realistic and a better predictive tool.
23 | P a g e
5.4.2 HYPOTHESIS 2
Tables 3 and 4 in the Appendix presents the results of the analysis on an aggregate level. The means and standard deviations of the utilities estimated on individual level are shown for each attribute and its respective levels. Additionally, the importance of each attributes is computed respective to the others. The results for both the first and second part of the survey are presented. Change in the aggregate utilities for the levels of the different attributes is present in the attributes related to extras. In the 1st part of the survey, representing 3 months before the purchase, point in time, the most preferred level of the attribute Extras was the third one followed by the first and second levels. In other words, having three months before the actual purchase, the respondents preferred having a wi-fi only rather than only GPS or the combination of both. At later point in time, however, just before the purchase, the respondents are more informed about the product and its different characteristic which causes a shift in their preferences. In the second phase the respondents’ preferences regarding the different combinations of extras have changed to the third level being most preferred again, but now the second level comes after that, followed by the first one.
Additionally, a more drastic change in preferences for the different attribute levels is present in the fitness extras. In the first phase of the survey, the respondents’ most preferred level is the second one followed by the first and the third one. Meaning that their first choice was a combination of heart-rate monitor and an accelerometer, followed by heart-rate monitor and a compass, and last but not least, accelerometer, pedometer and compass. When they received more information however, the respondents’ preferences changed in the following order : their most preferred level of fitness extras was now the first one, followed by the third and the second one.
24 | P a g e
For the purpose of examining the shift in attributes’ importances, computing them for the second part of the survey was done, disregarding battery life, so that the relative change, from the first part of the survey , can be examined. When it comes to the attributes’ importances relative to one another, their sequence did not change – the most important is price, followed by brand, extras, band and fitness extras. Despite of that, the percentages importance did change, meaning that a shift in preferences occurred. Additionally, when an extra attribute was added – the battery life, it did receive some importance – became next most important after the brand, and therefore, moved some attributes in a rank of importance which again is a useful to know for a manufacturer, for example. Having those two in mind, it can be concluded that the consumer preferences of the different products, which are a combination of different attribute levels, will change in time with the acquisition of additional information.
The attributes importances are also examined on an individual level. Plot 1 and 2 in the Appendix represent the individual importances of the attributes for the first and second phase of the survey, respectively. Again, a shift is evident, both in the means and in the ranges of the attribute imporances proving that they do change in time with the acquisition of additional information.
5.4.3 HYPOTHESIS 3
The first indication of increase in respondents’ certainty of choice is presented in Table 2 of the Appendix which contains the frequency of the selection variable. An increase in the number of 0% and 100% allocated probabilities is noticed from the first part of the survey (3 months before the purchase point in time) to the second one (exactly before the purchase). On one hand, having more 0s means that the respondents were deciding between less number of alternatives and therefore giving a percentage greater than 0 to, say, 2 out of the 4 alternatives. On the other hand, increase in the frequency of the 100% allocation of choice is a clear indication of decrease in uncertainty since the respondents knew with a 100% confidence what they want out of the choice set.
25 | P a g e
standard deviations are presented. There is an increase in the standard deviations but it is a slight and hard to notice one. For that reason, a standard deviation was computed for each attribute level on the individual coefficients of the respondents. The results are presented in table 5 of the Appendix. In all cases except for the price coefficients there is an increase in the standard deviation indicating a greater heterogeneity of preferences. The reason for the decrease in the standard deviation of the price coefficient might be that the respondents became more price sensitive ( as it is indicated by the coefficient in the aggregated results – it went down from negative 0.09 to negative 0.11) due to increase in knowledge of the product and had a certain maximum price level. Additionally, a t-test was performed to examine the significance of the difference in the standard deviations. The results from the test, presented below, indicate that the standard deviation of the estimated coefficients for the two phases of the survey are significantly different from each other which is, again, an indication that the heterogeneity of preferences changes over time.
26 | P a g e
6. DISCUSSION
Conjoint analysis has, for long being used, as a tool for predicting consumer behavior. Its foundations date back to 1964 (Rao & Vithala, 2014) and it has been developed ever since. Most recently, it has been argued that using stated choice probabilities as an input is a better representation of reality as compared to using only a single product choice because time and uncertainty cannot be accounted for. (Manski,1999) Additionally, providing a single choice means that the particular product is the best one, from consumer’s perspective, currently but it does not mean there are other possible end results.(Manski, 2004)
When it comes to the market of innovative products, more particularly the one for smart watches, there is a lot of consumer uncertainty in the product cause by the new purpose of the product, lack of information of usage, limited knowledge, etc. ( Having that in mind and the amount of information that we are exposed to every day, together with the pool of information that we have access to, it would be unreasonable not to account for that influence it can cause. When a consumer is being uncertain, he or she will surely search for information to increase their confidence in the product before they make a purchase. Therefore, time can be a huge factor when customers define their preferences.
This paper aims at investigating the presence of uncertainty in the smart watch product. Many say that the main confusion when defining one’s preference is caused by the fact that up until now, the watch was considered an accessory rather than functional product. The lack of information of previous usage adds up and the end result is low consumer confidence. In order to resolve this issue consumers are very likely to postpone their purchase in time for the purpose of acquiring additional information which will certainly lead to a better personal choice. Therefore, the main objective of the paper is to show that, indeed, uncertainty for that particular product is present and time has an effect on it and on the importance people allocate to the different attributes when evaluating it.
27 | P a g e
Kim & Krishnan, 2015). The results of the study are in line with the theory showing that, indeed, uncertainty is present in the market for smartwatches. Additionally, as suggested by Manski (1999), time is a factor to be considered since uncertainty in the product can be influenced by additional information acquired. This is exactly the case for the smartwatches as presented in the results – time or more precisely additional information, is shown to affect consumer’s preferences. In particular, as proposed by Oppewal et al. (2010), the order of importance that the consumers put to the different levels of the attributes changes with time. Last but not least, according to Huber et al.( 2002), increase in consumer certainty in their choice increases the heterogeneity in the sample because having exact preferences increases the differences in the different consumers’ opinion. Such is the case for the smart watches as well. The results indicate increase in the heterogeneity of preferences which means that the preferences of the different respondents shifted as soon as they broadened their knowledge about the product at stake.
In conclusion, uncertainty is present in the market for innovative product, more precisely in the market for smartwatches, and time is a big factor to be considered. If disregarded, the results can be a wrong predictor of behavior. Therefore, in a case of new product development, for example, the consumer preferences will be a bad predictor of behavior which will make the investment pointless and will eventually lead to a loss. In that sense, it is very important for managers to be very well acquainted with the market and understand its specificity before conducting a conjoint analysis. In a case of uncertainty being present, accounting for time is a must and a way to do it is using stated choice probabilities in the analysis.
7. LIMITATIONS
There are three main limitations of the study to be considered if future research on the topic is conducted.
28 | P a g e
Next, studying time as a moderator is done in single point in time by using a proxy. It is expected to reach more accurate conclusion if the time is accounted for in its real form. In other words a study, asking for the preference of the same respondents today and then surveying them again after 3 months have passed will include the real-life effect of time on uncertainty and product choice. Third, as evident in the results, the attribute price takes up great part of the product importance which can be due to the fact that, currently the participants in the survey are not very well acquainted with the smartwatch, therefore not confident enough and as a result more price sensitive. For that reason, excluding the price as a factor when creating the survey might be considered as an option.
Last but not least, separating the effect of changing uncertainty from the changing level of heterogeneity might be a difficult task but has to be considered. When the heterogeneity in the sample does not vary much the standard deviation of the individual preferences can be used as a measure of uncertainty. However, when this is not the case an issue because of the scaling might arise. (Louviere, 1988)
8. CONCLUSION
29 | P a g e
REFERENCES
AJZEN, I., AND M. FISHBEIN (1980): Understanding Attitudes and Predicting Social
Behavior. Englewood
Cliffs: Prentice-Hall
Brazell, Jeff D., Christopher G. Diener, Ekaterina Karniouchina, William L. Moore, Válerie Séverin, and Pierre-Francois Uldry. "The No-choice Option and Dual Response Choice
Designs." Marketing Letters Market Lett 17.4 (2006): 255-68. Web.
Chartrand, Tanya L. "The Role of Conscious Awareness in Consumer Behavior." Journal of
Consumer Psychology 15.3 (2005): 203-10
Delavande, Adeline. "Pill, Patch, Or Shot? Subjective Expectations And Birth Control Choice*." International Economic Review 49.3 (2008): 999-1042. Web
Desarbo, Wayne S., Venkatram Ramaswamy, and Steven H. Cohen. "Market Segmentation with Choice-based Conjoint Analysis." Marketing Letters 6.2 (1995): 137-47.
30 | P a g e
Eggers, F. [Felix] & Eggers, F. [Fabian] 2011. Where have all the flowers gone? Forecasting green trends in the automobile industry with a choice-based conjoint adoption model.
Technological Forecasting & Social Change, 78, pp. 51-62.
Green, Paul E., Abba M. Krieger, and Yoram Wind. "Thirty Years of Conjoint Analysis: Reflections and Prospects." Interfaces 31.3_supplement (2001): n. pag. Web.
Green, Paul E., and V. Srinivasan. “Conjoint Analysis in Consumer Research: Issues and Outlook”.Journal of Consumer Research 5.2 (1978): 103–123
Harmen Oppewal , Mark Morrison , Paul Wang , David Waller (2010), Preference Stability: Modeling how Consumer Preferences Shift after Receiving New Product Information,
in Stephane Hess , Andrew Daly (ed.)Choice Modelling: The the-art and The
State-of-practice (Default Book Series, Volume ) , pp.499 – 516
Hierarchical Bayes Conjoint Analysis: Recovery of Partworth Hetergeneity from Reduced
Experimental Designs." Hierarchical Bayes Conjoint Analysis: Recovery of Partworth Hetergeneity
from Reduced Experimental Designs. N.p., n.d. Web. 05 June 2016
Huber, J, Ariely, D., Fischer, G., 2002. Expressing preferences in a Principal-Agent task: A comparison of choice, rating and matching. Organizational Behavior and Human Decision
Processes, 87 (1), 66-90
Juster, T.(1966): “Consumer Buying Intentions and Purchase Probability: An Experiment in Survey Design,” Journal of the American Statistical Association, 61, 658–696.
Kim, Youngsoo, and Ramayya Krishnan. "On Product-Level Uncertainty and Online Purchase Behavior: An Empirical Analysis." Management Science 61.10 (2015): 2449-467. Web.
Littler, Dale, and Demetris Melanthiou. "Consumer Perceptions of Risk and Uncertainty and the Implications for Behaviour towards Innovative Retail Services: The Case of Internet
Banking." Journal of Retailing and Consumer Services 13.6 (2006): 431-43. Web.
Louviere, Jordan J.. “Conjoint Analysis Modelling of Stated Preferences: A Review of Theory, Methods, Recent Developments and External Validity”.Journal of Transport Economics and Policy 22.1 (1988): 93–119.
Malhotra, N.K., 1982. Information load and consumer decision making. Journal of Consumer
Research 8 (4) March, 419-430.
Manski, Charles F., Kenneth I. Wolpin, and Elke U. Weber. "Analysis of Choice Expectations in Incomplete Scenarios." Elicitation of Preferences(1999): 49-72
31 | P a g e
Peterson, Robert A., and Maria C. Merino. "Consumer Information Search Behavior and the Internet." Psychology and Marketing Psychol. Mark.20.2 (2003): 99-121. Web
Pogue, David. "Smart Watches Flunk Out." Scientific American. N.p., 20 May 2014. Web.
http://www.nature.com.proxy
ub.rug.nl/scientificamerican/journal/v310/n6/full/scientificamerican0614-37.html
Rao, Vithala R. Applied Conjoint Analysis. Berlin: Springer, 2014. Print.
Simmons, S. & Esser, M. 2000. Developing Business Solutions from Conjoint Analysis. Springer,
pp. 67-96
Vriens, Macro, Oppewal ,Harmen, and Wedel ,Michael. "Ratings-based versus Choice-based Latent Class Conjoint Models-an Empirical Comparison." Market Research Society. Journal of
32 | P a g e
APPENDIX
TABLE 1 – ATTRIBUTES AND LEVELS
Attributes Levels
Brand Apple Samsung Assus
Price 300 200 100
Band Leather Rubber Metal
Extras GPS GPS, Wi-Fi Wi-Fi Fitness extras Heart-rate monitor, Compass Heart-rate monitor, Accelerometer Accelerometer, Pedometer, Compass Battery life 8h 10h 15h
33 | P a g e
34 | P a g e
TABLE 3 – SURVEY 1st PART POPULATION MOMENTS
. b stdev Range Importance
35 | P a g e
TABLE 4 – SURVEY 2nd PART POPULATION MOMENTS
. b stdev Range Importance
No_choice -0.75 2.1 Brand Apple 0.75 2.1 1.57 6.25% Samsung -0.82 2.1 Assus 0.07 2.83 Price -0.11 1.8 22 87.61% Band type Leather 0.28 1.9 0.48 1.91% Rubber -0.2 1.9 Metal -0.08 2.69 Extras GPS -0.45 1.9 0.71 2.83% Wi-Fi 0.19 2 GPS and Wi-Fi 0.26 2.55 Fitness Extras Heart-rate monitor, Compass 0.13 1.9 0.35 1.39% Heart-rate monitor, Accelerometer -0.22 1.9 Accelerometer, Pedometer, Compass 0.09 2.69 Battery life 8h -0.53 1.9 10h -0.34 1.9 15h 0.87 2.69
TABLE 5 – STANDARD DEVIATION OF POPULATION MOMENTS
Attribute part 1 -Stdev(b) part 2 -Stdev(b)
No_choice 0.7729 1.2908
Brand
Apple 0.7302 1.6197
Samsung 0.4402 1.2255
37 | P a g e
38 | P a g e
39 | P a g e
PICTURE 1 – SURVEY SCREENSHOTS
40 | P a g e
