he Effect of Peer Specific Information on Consum
lysis of a more complex product, a health insurance
T
er Preferences
A conjoint ana
rick Jansema E nd Business University of Groningen ics a Faculty of Econom h T esis MSc Marketing n Ja uary 17, 2017 st Supervisor: Hans Risselada aarten J. Gijsenberg 1 2nd Supervisor: M Erick Jansema Grote Beerstraat 360 9742SM Groningen t.rug.nl Tel.: +31621202792 ‐Mail: E e.jansema.1@studen S tudent Number: 2779366Dedication
To my parents and girlfriend, for their unconditional support and encouragement.Acknowledgements
I would like to thank my supervisor Hans Risselada who offered inspiring, critical and above all helpful feedback. I would also like to thank my thesis group colleagues for the help they offered.Abstract
Introduction
In the Netherlands it is stated by law that consumers can only change their health insurance in January, excluding exceptions, and they do in large numbers. In 2016 more than one million consumers changed their health insurance (Vektis, 2016). However, before consumers actually change their insurance they actively search for information, cues and/or heuristics, for example on the internet, that might assist or facilitate them in making a decision (Todd & Benbasat, 1992). Peer influence is one of the heuristics contributing in this process according to Sundar, Xu & Oeldorf‐ Hirsch (2009). When consumers make product choices, they do not only consider the attributes of a product but also the preferences of other consumers, such as peers (Narayan, Vithala, & Saunders, 2011). Peer influence or social influence is defined as “a change in a person’s cognition, attitude or behavior which has its origin in another person or group” (Raven, 1964). Furthermore, several studies showed that the impact of a peer’s influence is indicated by the similarities between the consumer and the peer (Smith, Menon, & Sivakumar, 2005). The principle of homophily is a way to describe this phenomenon. “This principle structures ties of every type which results in homogeneous networks with regard to sociodemographic‐, behavioral‐, and intrapersonal characteristics”. “Homophily in race and ethnicity creates the strongest divides in our personal environments, with age, religion, education, occupation, and gender following in that order” (McPherson, Smith‐Lovin, & Cook, 2001). According to Lazarsfield & Merton (1954) and Lin (1982) a peer’s influence will be the greatest when the consumer and peer share these characteristics. These findings however, are based on researches with consumer goods as a focal product and at a product aggregate level.
and then choose the product with the greatest utility. This popular analysis assumes that the attribute preferences are independent of the choices of others. The research of Narayan et al. (2010) suggests otherwise. They found that consumers update their attribute preferences in a ‘Bayesian manner’. Which means that the consumer’s (revised) preference for an attribute is a ‘weighted’ average of her prior preference and the preference of the peer. Narayan et al. (2010) also found that the revision of attribute preferences, because of peer influence, varies across attributes. The influence is positive but it varies. However, peer influence in this study only took observational learning into account (Zhang, 2010). Consumers did not receive any other peer specific information. The information sharing part, Zhang (2010) suggests, has been disregarded. Until now because we are about to fill this gap in research. In this study we proceed with the recommendation of Narayan et. al (2011) to extent prior research with the incorporation of more peer specific information in the decision making process, at an attribute level. In which we have chosen to share information about the peer’s characteristics derived from the principle of homophily. Thereby, in addition, this research is not about consumer goods. In contrast to most other studies, a more complex and important product is used as the focal product, health insurances. Having a health insurance is legally required, above the age of 18 (individually), and signing a contract with a company is for at least one year. Every January consumers have the opportunity to switch, after this month people have to wait for at least another year. The real complexity in choosing a health insurance is in the fact that this insurance is an insurance against the risk of incurring medical expenses. Something people don’t know in advance, which makes it very difficult but also very important to make a good and deliberate decision.
more complex, in comparison with the Narayan et al. (2010) study who used E‐readers and cell phones as their focal product, the consumers probably put more emphasis on the ‘price’ people might have to pay. This is supported by a research of Tung‐Zong and Wildt (1994) who found that indeed consumers rely more on price in making judgments and decisions for complex products. Besides this, a second addition to prior research is the fact that more peer specific information (based on the characteristics providing the greatest differences and similarities between people) is included in the study. Since Lin (1982) found that sharing characteristics leads to greater peer influence, the results in this study will probably illustrate an even greater effect of peer influence, in comparison with prior studies which did not account for peer specific information (i.e. Narayan et al. 2010). The research question arising from the complexity of the product and the supposed influence of a peer is the following: How does peer specific information change the attribute preferences of a more complex product, a health insurance?
A conjoint experimental design is chosen to estimate the effect of a peer in the decision making process. Louviere, Hensherr & Swalt (2000) introduced the choice‐based conjoint (CBC) approach. In which consumers choose the most preferred product derived from a set of alternatives. This approach makes the task of deciding which product is the most preferred very easy and effective in terms of determining consumers’ preferences.
consumer decision attribute provides information about what random people and people like you did. So a distinction has been made between the people that look like you and the people who don’t. In this way peer influence is incorporated. During the third part the actual conjoint will be conducted which will result in consumers attribute preferences.
As suggested by Homans (1950), Lin (1982) and Lewis, Gonzalez & Kaufman (2012) the results correspond to the idea of greater peer influence due to more shared characteristics. The attribute evaluated as the most important is the amount of own risk and the attribute indicating peer influence (consumer decision) was judged as the least important. Knowing that consumers do not judge peer influence as the most important but prefer the alternative in which the amount of own risk was the lowest is very interesting in terms of practical implications. In comparison with the consumer goods a peer seems to have less influence in the decision making process. The complexity of a product is thus of decisive importance. That is what we contribute to theory but it might also have some consequences in a more practical perspective. Marketers, i.e., should put less emphasis on what others did and focu o the product and the importance of the decision itself. s n
The remainder of this paper is organized as follows: in the next section the conceptual model will be discussed, together with the formulated hypothesis. After that the methodology and data collection will be described in more detail. Next is the results section and the empirical application. Finally, we conclude with the discussion and implications for marketing practices and
esearch. r
Conceptual framework
In this study consumers choice for a health care insurance is considered. A distinction has been made between four product attributes. The insurances do not differ apart from these attributes. In the Netherlands 90% of the market is owned by four major companies, Zilveren Kruis, VGZ, CZ and Menzis (Vektis, 2016). However to overcome brand preferences up forehand we do not include
rand a
b s an attribute. Consumers income is included in the framework as a moderator.
The first attribute is consumer decision in which the influence of a peer is included. A distinction in attribute level has been made between ‘people like you’, based on the characteristics providing the greatest differences and similarities between people. ‘Other people’, in which no further information has been shared apart from the actual decision (only observational learning). To compare and learn whether adding peer specific information really leads to greater peer influence. The third attribute level, ‘no information available’ has been included to actually find out if a peer and peer specific information add ‘value’ in a consumers decision making process.
Due to the inter‐attribute variation because of peer influence, as suggested by Kahn and Meyer (1991); Green and Krieger (1995) and Narayan et al. (2010), the moderating effect of this attribute on the other three attributes will also be determined.
contribution, and take more risk with respect to the amount of own risk they might have to pay. This trend is very nteresting and we are curious whi ether a peer could influence this.
The third and fourth attribute are part of the additional insurances in the Netherlands. Consumers are not required to take one of the additional insurances, however 84% of the insured people do (Vektis, 2016). But why did we pick these additives (the amount of physiotherapy treatments and the amount of dentist fee people have to pay) as the third and fourth attribute? Looking at the distribution of the compensated healthcare costs the results show that the dentist (44,5%) and paramedic treatment (physiotherapy) (27,2%) are by far the best covered. Which is very important for the people, to avoid unforeseen costs. They also found that these two ‘attributes’ are the most important reason for people to take an additional insurance (Vektis, 2016). These four attributes form the base of this research.
Income is included in the survey as a socio‐economic variable. Five income levels are determined based on the income distribution in the Netherlands (CBS, 2014). A study of Paridon, Carraher & Carraher (2006) found that consumers preferences for a multi‐attribute product are influenced by income. Which also makes sense in this study, and especially since the costs involving medical care are pretty high. Therefore, income is included in the model as a moderator. The
Hypothesis
The effect of peer influence is very clear and well documented (Craig, & Bush, 2000; Dahl, Fischer, Johar, & Morwitz, 1982). Childers and Rao (1992) for instance showed that peer influence is positively related to consumer product choice. As well as Wang, Yu and Wei (2012) who showed in a more recent study that a peer even has influence on the purchase intention of a consumer through social media. Therefore, a logical hypothesis is the following. H1: Peer influence will have a positive impact on consumer preferences. Positive in this context means that the revised preference of a consumer is more in line with the preference of the peer, at an aggregate level. In the conjoint a peer’s influence is indicated as the attribute named ´consumer decision (others)’.
H2: The amount of own risk consumers pay will be judged as the most important.
What Vektis (2016) also reported is a trend in the amount of own risk people might have to pay. As mentioned before, the amount of people paying a higher and voluntary own risk increased with 6%, from 6% up to 12%. So instead of paying the €385 (basic) people consider and actually choose to pay a higher own risk, i.e., €635, and in return consumers pay less each month. Which is quite a risk. However, it is a very interesting trend and we therefore hypothesize the following.
H2a: Consumers will have a greater preference for paying voluntary own risk.
The third and fourth attribute are part of the additional or supplementary insurances. Consumers are not required to choose one of the additional insurances however, 84% of the people do so (Vektis, 2016). A distinction within the additional insurance can be made between i.e., dentist costs, physiotherapy treatments, alternative care and pharmaceutical care. Almost 75% of the people who have an additional insurance are covered for an amount of physiotherapy treatments and dentist expenses (Vektis, 2016). With slightly more people having an additional insurance for the dentist.
H3: The amount of physiotherapy treatments will be the least important.
H4: The mount of dentist xp nses covered will be the second most important.
Least and second most important in comparison with the amount of own risk, amount of hysiotherapy treatments covered and the amount of dentist fee people have to pay.
a e e
p
The people having a higher income are among the bigger risk takers, according to MacCrimmon and Wehrung (1990). So having a higher income makes people 1) less price sensitive, 2) less sensitive about what others think and do and 3) more risk taking. Therefore the following hypothesis are determined.
H5: Higher income leads to less price sensitivity
H5a: Higher income is negatively re ated to peer’s influence
Less price sensitivity is an indicator for not avoiding the alternatives including more expensive medical expenses. Negatively related in this case is an indicator for putting less emphasis on a
eer’s decision.
l
Methodology
Study design
Health insurances are chosen as the focal product in this study. The main purpose of the study is to investigate in what extent people are affected by the choice other people make. In which the information is illustrated in the conjoint, about what other people chose, differs in three ways, see table 1. Furthermore, consumer’s attribute preferences will be determined. The design containing four attributes with three levels is shown in table 1.
Table 1: Attributes and levels
Attribute Levels Specification
Consumer decisio thers) n (made by o (1) People like you chose this people chose this op ormation available n optio tion (2) Other (3) No inf Part‐worth Own risk (1) €385 (2) €635 (3) €885 Part‐worth Physiotherapy treatments (1) 10 (2) 25 (3) 40 Part‐worth Dentist fee (1) €250 (2) €500 (3) €750 Part‐worth
The insurances do not differ apart from the attributes that are listed. Three out of the four attributes are determined based on a research of Vektis (2016), in which these attributes turned out to be the most important to consumers. The other attribute (consumer decision) is based on theory explained in the previous sections. In the survey each attribute was explained extensively and a tooltip was included for those consumers who wanted to read the information once again.
he information illustrated in the survey is shown in table 2. T
Table 2: Tooltips attributes
Attribute Tooltip/ explanation Consumer
decision
“What did other people choose. Differs in three levels namely; other people,
people that look like you (defined using your answers in the previous questions) nd no information available.”
a
Own risk “Legally required amount of own risk, €385. The own risk is the amount you have
to pay before the insurance company covers the medical costs. Premium each month decreases with €10 (for €625) and with €20 (for €885).”
Physiotherapy treatments
“Amount of treatments covered a year. Premium each month increases with €10 (for 10), with €20 (for 25) and with €30 (for 40).”
Dentist fee “Covered until a certain amount of money, illustrated by the different levels. Premium each month increases with €10 (for €250), with €20 (for €500) and with €30 (for €750).”
Research design
The amount of possible health insurance combinations is 81 (3*3*3*3) and three stimuli per set, which results in 85.320 potential choice sets (81!/(81‐3)!*3!), will be used. The number of choice sets is reduced to twelve by using a fully‐randomized, with minimal overlap, choice design containing three health insurance alternatives. According to Eggers (2014) twelve is a good number to obtain as many information as possible and to avoid respondents from not completing the survey. A no‐choice option is not included in the survey since Dutch citizens are legally required to have a health insurance. A control question in which is asked to perform a certain task, for example: hit button 7, is not included. However, something else is included to exclude respondents that did not read carefully. At the bottom of the ‘information/ processing’ page the following sentence is written: “Due to the processing it might take some seconds before the next questions show up, good uck!”. It only took 4 seconds of patience but the completion rate still dropped with 10.1%. l
Data collection
The survey was administered online to residents of the Netherlands, also foreign people living and/ or working in the Netherlands are included. The respondents are aged between 18 and 67.Model diagnostics
To test whether the estimated models are significantly different from the null models, likelihood ratio tests are performed using a chi‐squared test statistic, which is chi‐square distributed with
L(0) is degrees of freedom (df) equaling the number of parameters: X2 = ‐2(LL(0) – LL(β)), where L
the log like ihood of the null model and LL(β ) is the log ikelihood of the estimated model. l l
To assess whether the estimated model fits the data well, McFadden’s adjusted R2 is
determined using the following formula: R2adj. = 1 – (LL(β) – npar)/LL(0), where LL(0) and LL(β)
are defined as mentioned above. A number between 0,2 and 0,4 can be considered acceptable (Eggers, Hauser, & Selove, 2016). To select a model, log likelihood information criteria are used. More specific, the model with the lowest Consistent Akaike Information Criteria(CAIC) will be used. BIC and CAIC are preferred in large sample sizes and give higher penalties for complexity (i.e. more latent classes). CAIC is defined as: ‐2 * LL(β) + (ln(N) +1) * npar, where npar is the amount of parameters and N is the sample size. To assess how well the model predicts respondents’ actual choices, hit rates are calculated, using the sum of the estimated observations and dividing it by the total observations. So how many of the estimated observations were actually right. The higher the outcome, the better the model.
By looking at the Classification Error the latent class classification will be evaluated. The Classification Error averages the minimum posterior class membership probabilities across
onsumers, the lower the error the better the model. c
Results
Sample statistics
The study sample includes Dutch residents above the age of 18, with a maximum age of 67 and a mean of 31. The total number of respondents is 207; 112 were completed, yielding a completion rate of 54%. In which the most people dropped out after or at the first page (28,1%). Most of the respondents were male (64,3%); further respondent characteristics are shown in table 3. Table 3: SaVariablemple Statistics Clas isif cation Mean(S.D.) Sample (%)
Gender (0) Male
(1) Female
64,3%
35,7%
Age In year (18+) s 31,3 (11,9)
Main model
The first model that has been estimated is the main model. The model statistics can be found in table 4. If we compare the log‐likelihood (LL) of this model with the null model (‐1476,5349) a decrease of the LL can be found.
Table 4: Model statistics
Model LL BIC CAIC Npar Class. Error R2 adj. Hit Rate
1‐Class ‐1214,395 2466,538 2474,5318 8 0,000 0,1721 61,16% To determine if both models are significantly different the likelihood ratio test has been conducted. The chi‐square test statistic turned out to be 524,2798. The critical value (α=5%) for a chi‐squared distribution with 8 degrees of freedom is 15,507. Hence, we reject the null hypothesis and conclude that the estimated model parameters are significantly different from zero.
Parameter interpretation
First of all we start with looking at the significance of the four attributes. The attributes ‘Own risk’ and ‘Dentist’ turned out to be highly significant (p<0,001), as shown in table 6. The attribute ‘Consumer decision’ is significant at 5% level. The amount of physiotherapy treatments turned out to be insignificant (p>0,05). We therefore do not interpret these part worth’s, they do not significantly differ from zero. If we then look at the part worth’s of the consumer decision attribute we are able to conclude, although the part worth’s are very small, that consumers prefer the alternative in which the people that look like them is included. Therefore, hypothesis 1, “peerinfluence will have a positive impact on consumer preferences”, is accepted. Furthermore, by looking
very close to zero. The same more or less holds for the covered amount of dentist treatments. eople tend to prefer the option with the lowest costs.
P
The relative attribute importance is calculated as a percentage of the largest difference between the attribute parameters, the results are illustrated in table 5.
Table 5: Relative attribut
portan
e importance
Relative Im ce In percentage Consumer decision 8,15% Own risk 64,07% Physiotherapy 6,30% Dentist 21,49% The relative importance of the attribute ‘consumer decision’ turned out to be 8,15%. So although the attribute has a positive effect on consumer preferences, whereby H1 is accepted, the relative
importance is very low. The most important attribute, by far, is the amount of own risk people have to pay, that is why hypothesis 2 is accepted. The third hypothesis was that the attribute ‘dentist fee’ will be judged as the second most important. The results confirm these findings, and therefore hypothesis 3 is accepted as well. The amount of physiotherapy treatments covered turned out to be insignificant and is also evaluated as the least important, so hypothesis 4 is accepted.
Table 6 – attribute parameters main model
Attributes Class1 Wald p‐value Mean Std.Dev.
Dentist 1 0,2004 44,0426 2,70E‐10 0,2004 0 2 0,0939 0,0939 0 3 ‐0,2943 ‐0,2943 0
Model including interaction effects
The second model that has been estimated is an extension of the main model. The interaction effects are included. The model statistics are shown in table 7, the statistics of the first model are included as well. The second model seems to decline in model fit. The information criteria increase and the R2 adjusted decreases, same holds for the hit rate. Conducting a likelihood ratio test results
(chi‐sq= 31,914, with 36,415 as the cutoff point with 24 degrees of freedom) in an insignificant outcome. Thus, both models do not really differ. However to determine the interaction effects we proceed with this model.
Table 7: Model statistics model 1 and 2
Model LL BIC CAIC Npar Class. Error R2 adj. Hit Rate
1‐Class ‐1214,395 2466,538 2474,5318 8 0,000 0,1721 61,16% 1 i ‐Class + nteractions ‐1198,438 2547,485 2579,8680 32 0,000 0,1667 61,10%
As mentioned before, to include interaction effects in the model new variables indicating these interactions have to be determined. Each new variable indicating interaction effects of two variables is the outcome of both attribute (levels) multiplied with each other. So, income has been multiplied by each level of the four product attributes (consumer decision, own risk, physiotherapy and dentist) and each level of consumer decision has been multiplied by each level of the other three product attributes. Which resulted in eight new income variables and twelve new consumer
ecision variables, the part worth’s are shown in appendix A. d
Parameter interpretation
By looking at the significance level of, first, the variables including an interaction effect between consumer decision and the other three attributes, none of the variables turned out to be significant, also illustrated in appendix A. Therefore hypothesis 1a is rejected. On the other hand, the variables including income do show some significant interaction effects. One of the attributes interacting with income is the amount of own risk, the part worth’s are shown in table 8. Table 8: inte ibute raction effect of income with o n risk w
Attr Class 1 Pvalue
Own risk 385 0,9056 4,4e^‐57 635 ‐0,0879 885 ‐0,8177 Own risk_*_income_1 ‐0,0830 0,021 Own risk_*_income_2 ‐0,0429 0,32 The effect of an alternative including an amount of own risk of €385 decreases by ‐0,0830 for every unit of income. So the higher someone’s income the lower the preference for the lowest amount of own risk (€385). In perspective, consumers having the highest income still prefer the option in which the amount of own risk is the lowest, since the decrease is very small. The other interactions are insignificant. Therefore, H5 will only partially be accepted. A negative interaction effect between
a higher income and peer influence, as hypothesized in hypothesis 5a, has not been found, the hypothesis is rejected.
Segmentation
Latent Gold is used to estimate seven models, first without the interactions, to define the number of segments. Three out of these seven are illustrated below, the others are in appendix B. The model with 4‐classes is considered the most appropriate since this model has the lowest CAIC. Table 9: Model statistics latent class models5‐Class ‐921,4645 2050,5430 2094,5430 44 0,0420 0,3461 76,04% 6‐Class ‐899,3166 2048,7137 2101,7137 53 0,0349 0,3550 77,45%
The chi‐square test statistic turned out to be very large and therefore the null hypothesis is rejected, this model differs from the previous (two) models. The respondents are very nicely divided across the four classes, even though the first class is the biggest in size, illustrated in appendix C. The attribute parameters for each class are summarized in table 11, on the next page. By adding covariates the classification error of the model can be improved and these covariates might explain the differences in segments. The own‐risk‐, physiotherapy‐ and dentist parameters turned out to be highly significant (p<0,001). The consumer decision parameter is significant on a 5% level. The four different classes significantly differ in only three attributes. Consumer decisions turned out to be insignificant (p>0,1). Some interesting differences between classes can be identified. For instance by looking at the attribute levels of the amount of own risk. The classes significantly differ since the p‐value, the second p‐value column in table 11 indicating this, is very small (p<0,001). The results show that the first three classes put the most weight on the lowest amount of own risk, see appendix C. The fourth class, on the other hand, completely differs from the other three classes. Consumers in that class prefer the highest amount of own risk and do not like the option in which the lowest amount of own risk is presented at all.
Attribute importance
Looking at how preferences differ across classes, the relative attribute importance is calculated as a percentage of the largest difference between the attribute parameters, shown in table 10. Table 10: Relative attribuAttribute teClass1 importance Class2 Class3 Class4 across cla es ss Consumer decision 8,29% 15,33% 2,63% 12,97%
Own Risk 67,75% 35,96% 34,75% 29,42%
Physiotherapy 10,68% 28,32% 10,23% 26,36%
Dentist 13,27% 20,39% 52,38% 31,25%
Class size 38,58% 22,16% 21,84% 17,41%
The first class is characterized by an overwhelming preference for the own risk attribute. The other attributes are not that important. The second class also put the most emphasis on the amount of own risk, however they also take the other attributes into account and the relative attribute importance is a bit more varied. The third class doesn’t care about what others or others like themselves did and put the most emphasis on the dentist attribute. The fourth and last class shows
ore variation, however the ‘dentist’ attribute is the most important. m
Table 11: Attribute parameters 4class model
Attributes Class1 Class2 Class3 Class4 Wald p‐value Wald(=) p‐value Mean Std.Dev.
Covariates
By adding covariates the classification of a model can be improved. To test whether there are covariates explaining the latent classes, variables containing consumer information (demographics) are included. However, each covariate turned out to be insignificant as shown in table 12.
Table 12: Significance Table 13: Significance Covariates all at once Covariates one by one
Covariate Pvalue Gender 0,31 Age 0,078 Education 0,90 Occupation 0,99 Income 0,65 Religion 0,84 When testing for multicollinearity between these variables, it appears that almost every variable strongly correlates with the other variables. Which might explain the previous outcomes. Religion is the only variable not correlating with two or more other variables as illustrated in appendix D. Therefore to test, once again, if one of the covariates significantly contributed in explaining the classes, every variable is included as a covariate apart from each other, shown in table 13. However, again none of the covariates turned out to be significant. Which means that the demographics we asked for do not seem to explain the latent classes. A lear explanation about how the consumers look like cannot be given. Covariate Pvalue Gender 0,73 Age 0,32 Education 0,66 Occupation 0,81 Income 0,43 Religion 0,82 c A fourth model has been estimated as well, including every possible interaction effect and Latent Gold, again, was used to determine latent classes. However, this model turned out to not significantly differ from the previous discussed model. The chi‐squared test statistic turned out
o be 13,1524, with 63 degrees of freedom. The model statistics are illustrated in appendix E. t
Discussion
To investigate in what extent consumers are affected by the choice of others a conjoint analysis is conducted with health insurances as the focal product. In which four product attributes are determined with three levels each: consumer decision, amount of own risk, amount of dentist fee covered and the amount of physiotherapy treatments. Prior studies (Narayan, Vithala, & Saunders, 2011) showed that consumers do not only consider product attributes but also (other) consumers’ preferences when making product decisions. By including peer specific information we extent prior research. Homophily in ethnicity, age, religion, education, occupation and gender create the biggest divides in our personal environment and were therefore used to determine the people that look like you. The hypothesis was that the people that looked like you would have the biggest impact on consumers preferences. The results supported these findings. However, on the other hand the relative importance of this attribute turned out to be very low. The amount of own risk turned out to be, by far, the most important followed by the amount of dentist fee. So the complexity and importance of a product seems to affect the influence of others and others that look like you on consumers decision. Hence, an answer to the main question is that peer specific information does influence consumers attribute preferences however relative to the other attributes is not a very important attribute.
By looking at the research some things stand out. In the first case, a small completion rate of the survey has been measured, which was only 54%. We believe that, besides the already mentioned explanations, this is due to the fact that the survey was in English and that the majority who participated in the survey was Dutch. Therefore, we would consider doing a future researc nh i Du ch. t
to the fact Kahn and Meyer used a consumer good as the focal product but could also come from the fact that consumers did not feel some kind of uncertainty among the product attributes. Another explanation could be that the measured resistance to peer influence comes from the average age of the respondents. Steinberg & Monohan (2007) i.e. found that between the age of 18 and 30 people develop the capacity to stand up for what they believe and resist peer pressures. Since the respondents in this research turned out to be quite young, this could be the case here.
Another interesting result was that income only had a significant interaction effect with one of the own risk attribute levels (€385), in which we accepted H5. However, we did not find
any results suggesting a negative relation between the amount of income and peer influence. As we hypothesized in H5a. A possible explanation for this could be that due to the fact that the
relative importance of the attribute, indicating peer influence, was very low. So low that having a higher income does not really effect this attribute. Another explanation could possibly be that the amount of people having a higher income was very low, i.e. only 8% of the people in this
e
research earns mor than €55.000,‐.
To identify different latent preference classes Latent Class models were used, four different classes were determined. However none of the demographics, asked for in the first part of the survey, could explain the differences in classes. So there are latent classes but the demographics do not seem to explain these classes. A clear explanation on how the consumers look like therefore cannot be given. One potential reason for this finding might come from data limitations, i.e., the sample did not contain enough cases to distinguish real differen es in classes. c
The results, however, showed some really interesting findings. The fact that a peer’s influence is of minimum value, in a consumers product decision process in which the product is more complex, makes future research in this type of products very interesting and worth the effort.
Appendices
Appendix A Main model, parameter estimates
Model for Choices Class1 Overall R² 0,2104 0,2104 R²(0) 0,2106 0,2106Attributes Class1 Wald p‐value Mean Std.Dev.
‐0,0679 2,7643 0,096 ‐0,0679 0 Income_Dentist_250 ‐0,0225 0,3385 0,56 ‐0,0225 0 Income_Dentist1_500 0,0538 1,8484 0,17 0,0538 0 CD_A_OR_A ‐0,0784 0,7183 0,4 ‐0,0784 0 CD_A_OR_B 0,0616 0,5014 0,48 0,0616 0 CD_B_OR_A 0,0004 0 1 0,0004 0 CD_B_OR_B 0,0437 0,2484 0,62 0,0437 0 CD_A_OR_FYSIO_A ‐0,062 0,5019 0,48 ‐0,062 0 CD_A_OR_FYSIO_B ‐0,0795 0,8126 0,37 ‐0,0795 0 CD_B_OR_FYSIO_A 0,009 0,0106 0,92 0,009 0 CD_B_OR_FYSIO_B ‐0,0721 0,667 0,41 ‐0,0721 0 CD_A_OR_DENTIST_A 0,1159 1,7864 0,18 0,1159 0 CD_A_OR_DENTIST_B ‐0,1259 2,013 0,16 ‐0,1259 0 CD_B_OR_DENTIST_A 0,0172 0,0389 0,84 0,0172 0 CD_B_OR_DENTIST_B 0,1145 1,7368 0,19 0,1145 0
Appendix B Latent classes
Segments LL BIC(LL) CAIC(LL) Npar pvalue Class. error Model 1 1‐class ‐1214,3950 2466,5381 2474,5381 8 5,3e‐437 0,000 Model 2 2‐class ‐1054,9329 2190,0804 2207,0804 17 7,5e‐377 0,0230 Model 3 3‐class ‐991,7095 2106,1000 2132,1000 26 1,1e‐356 0,0324 Model 4 4‐class ‐943,2627 2051,6729 2086,6729 35 9,7e‐343 0,0485 Model 5 5‐class ‐921,8782 2051,3703 2095,3703 44 3,5e‐340 0,0467 Model 6 6‐class ‐899,3166 2048,7137 2101,7137 53 2,6 ‐337 e 0,0349
Model 7 7‐class ‐883,1952 2058,9374 2120,9374 62 2,3e‐337 0,0321
Model 8 8‐class ‐867,9592 2070,9318 2141,9318 71 4,2e‐338 0,0409
A
ppendix C Class sizes
Class1 Class2 Class3 Class4
Appendix D Correlation check
Correlations
A
ppendix E ‐ Parameter estimates, interactions and latent classes
Model for Choices
Class1 Class2 Class3 Overall
R² 0,2853 0,7738 0,3698 0,4751
R²(0) 0,2853 0,774 0,3706 0,4754
Attributes Class1 Class2 Class3 Wald p‐value Wald(=) p‐value Mean Std.Dev.
‐ 0,2122 0,0943 0,1338 10,0451 0,018 9,7273 0,0077 ‐0,008 0,1583 Income_Dentist_250 ‐ 0,2668 ‐1,2666 0,4272 54,2426 1,00E‐ 11 54,1602 1,70E‐ 12 ‐0,3763 0,6757 Income_Dentist1_500 0,0312 0,2032 ‐0,0077 0,7797 0,85 0,6249 0,73 0,0746 0,0899 CD_A_OR_A ‐ 0,3528 1,4243 0,0267 15,7659 0,0013 15,5156 0,00043 0,3345 0,7662 CD_A_OR_B 0,1649 0,7792 ‐0,0687 4,3561 0,23 3,2269 0,2 0,2911 0,3494 CD_B_OR_A 0,2941 ‐0,6008 ‐0,3111 8,0893 0,044 8,0759 0,018 ‐0,1785 0,3822 CD_B_OR_B ‐0,034 ‐0,8376 0,2941 6,6715 0,083 6,6511 0,036 ‐0,1922 0,4642 CD_A_OR_FYSIO_A ‐ 0,0855 ‐0,1277 ‐0,181 1,3736 0,71 0,144 0,93 ‐0,1283 0,0392 CD_A_OR_FYSIO_B 0,0876 ‐0,5631 0,0019 3,2089 0,36 3,1664 0,21 ‐0,1478 0,288 CD_B_OR_FYSIO_A ‐ 0,0057 1,0524 ‐0,1059 8,5752 0,036 8,2744 0,016 0,3039 0,5169 CD_B_OR_FYSIO_B ‐0,197 ‐0,2027 ‐0,171 2,8906 0,41 0,0128 0,99 ‐0,1909 0,0134 CD_A_OR_DENTIST_A 0,0925 ‐0,5922 0,065 3,4081 0,33 3,4054 0,18 ‐0,1361 0,3142 CD_A_OR_DENTIST_B ‐ 0,2291 ‐0,1054 0,0046 1,9536 0,58 0,7591 0,68 ‐0,1179 0,0962 CD_B_OR_DENTIST_A 0,1778 0,4504 ‐0,0335 3,093 0,38 1,6784 0,43 0,2009 0,1924 CD_B_OR_DENTIST_B 0,1751 ‐0,3901 0,2125 4,1693 0,24 3,2191 0,2 0,0048 0,2723 Model for Classes
Intercept Class1 Class2 Class3 Wald p‐value
0,1163 ‐0,0324 ‐0,0839 0,7406 0,69
Study Design
Conjoint, with health insurance as the focal product
12 choice sets with each 3 alternatives
Each alternative consists out of 4 attributes and 3 attribute levels
Study Design
#Attribute 2
Own Risk
€385
€635
€885
Physiotherapy
10
25
40
Dentist
€250
€500
€750
#Attribute 3
#Attribute 4
Vektis is a company that gathers
and analyzes information about
health care and health services
in the Netherlands.
The amount of physiotherapy
treatments and the amount of
Dentist fee covered are
additional insurances.
Study Design
Results
Model 1: Main effects
Attributes Class1 Wald pvalue Mean Std.Dev.
Results
Model 2: Main effects + interactions
Decline in model fit?
Likelihood ratio test Æ insignificant
Only one interaction turned out to be
significant
Model LL BIC CAIC Npar Class. Error R2adj. Hit Rate
1‐Class ‐1214,395 2466,538 2474,5318 8 0,000 0,1721 61,16%
1‐Class + interactions ‐1198,438 2547,485 2579,8680 32 0,000 0,1667 61,10%
Attribute Class 1 Pvalue
Results
Model 3: Latent Classes
4 class model based on minimizing
CAIC
Model LL BIC CAIC Npar Class. Error R2adj. Hit Rate
4‐Class ‐943,2627 2051,6729 2086,6729 35 0,0485 0,3374 74,85%
5‐Class ‐921,4645 2050,5430 2094,5430 44 0,0420 0,3461 76,04%
6‐Class ‐899,3166 2048,7137 2101,7137 53 0,0349 0,3550 77,45%
Attribute Class1 Class2 Class3 Class4