Preference Variation: Where Does Health Risk Attitude Come Into the Equation?

Preference-Based Assessments


Introduction
Decisions about health often involve risk, but at the same time it must be recognized that different decision makers interpret and value risk information differently. 1,2For example, some people prefer a small chance of full recovery, over a more certain but moderately good outcome, and others vice versa.Examples such as these portray why patients' decisions sometimes conflict with physicians expectations or advice. 3The psychological construct that describes how people make decisions under uncertainty is called risk attitude, 4 which has been shown to affect health policy decisions on an aggregate level when incorporated in analyses. 5,6lso on an individual level, attitude toward health-specific risks contributes to heterogeneity in health preferences and behavior. 7,8nsights into the relationship between risk attitude and preference heterogeneity in health can improve the accuracy with which uptake and adherence to treatment are predicted. 9,10In addition, they can be informative when alternatives largely vary in terms of benefit-risk or when treatment outcomes are highly uncertain, 11,12 for example, by providing information tailored to varying health risk attitudes in clinical practice or for patient subgroup considerations in benefit-risk assessments.Although researchers generally agree that risk attitude plays a role in healthcare decision making, there is no consensus on how to operationalize it. 13isk attitude is often domain specific 14,15 ; it covers both risk perception and risk-taking behavior, although these can be conflicting concepts, 15,16 and risk attitudes as measured in surveys do not always translate to the real world. 17,18As such, studying the relationship between health-risk attitude and health preferences is rather complex.
Health preferences are often elicited via discrete choice experiments (DCEs), a method to quantify preferences for alternative health interventions by letting respondents repeatedly trade-off alternatives that are described using a variety of attributes and levels. 190][21] Many of the health interventions described in DCEs include one or more risk attributes.Harrison et al 22 systematically reviewed DCEs with a risk attribute in the field of health and found that few of these DCEs incorporated individual characteristics underlying risk preferences.Only Tsuge et al 23 elicited subjective risk perception and found that it influenced willingness to pay in their DCE.None of the other identified studies elicited risk attitude; instead they derived it from responses to the choice tasks and used this information in their statistical analysis.In line with the increasing use of and demand for analyzing heterogeneity in DCEs, 19,24,25 Russo et al 13 systematically reviewed individual characteristics underlying the decision-making process and their relation to preference heterogeneity.They identified risk attitude as 1 important, yet marginally studied, factor relating to preference heterogeneity; experts agreed with this assessment. 12Furthermore, studies that assess risk in health-related DCEs focus on risk communication rather than risk attitude. 26,27Although individual characteristics such as health numeracy, health literacy, and decision-making style (also identified by Russo et al 12 ) are increasingly found to be related to preference heterogeneity, 28,29 the complexities associated with the operationalization of risk attitude hamper studying the relationship between risk attitude and preference heterogeneity. 13herefore, the purpose of this study is to determine whether and how health-risk attitude and heterogeneity of health preferences are related by means of 3 case studies, using the relatively new 13-item Health-Risk Attitude Scale (HRAS-13), which aims to overcome some of the operational complexities. 16To assess the relationship between the HRAS-13 and heterogeneity, we studied 2 types of heterogeneity, namely, (1) stochastic class assignment and (2) systematic preference heterogeneity.

Case Studies
To study the association between health-risk attitude and heterogeneity of preferences, we selected 3 DCE case studies in the health domain that had at least 1 attribute that implicitly or explicitly concerned risks and for which we could gain the authors' consent to share the data for this purpose.The studies differed in terms of their topic, country, study population, number of respondents, and their DCE design leading to an increased generalizability of the results.1][32] Attributes and levels were selected based on literature reviews, focus groups, and interviews; these are presented in Table 2.The first case study concerned the treatment preferences of patients with multiple sclerosis (MS) in The Netherlands, France, and the UK. 30 Inclusion criteria were the following: aged 18 years or older, diagnosis of MS, and living in one of these 3 European countries.Respondents were recruited online via the commercial survey sampling company Survey Engine (N = 753).Three of 4 attributes were explicitly described as risks to survey respondents.The second study analyzed preferences regarding antibiotics usage in Sweden. 31An online sample of respondents between 18 and 65 years old was recruited from the Swedish general public (N = 378).Respondents were recruited online via Dynata, a commercial survey sample provider.Three of 5 attributes concerned risk, 2 of them in percent and the third in words.The third study concerned care for hip and knee osteoarthritis (HKOA) in The Netherlands. 32Respondents aged 45 years and older with knee or hip osteoarthritis were recruited online, also via Dynata (N = 648).In contrast to the other 2 studies, none of the attributes were explicitly described to respondents as being related to risks.Nevertheless, based on the relationship between time preference and risk aversion, "waiting time in weeks" was classified as a risk attribute. 33,34The number and type of professionals involved were also classified as a risk attribute because health anxiety increases the belief that specialist referral is needed, and health anxiety was found to be driven by risk aversion. 35,36[32]

DCE Design and Questionnaire
In all studies, a Bayesian heterogeneous DCE design was created using Ngene ChoiceMetrics software 37 to maximize D-efficiency.Initial priors were based on literature, focus groups, and interviews with experts or members of the study population in the prepiloting phase.Based on the results of a standard multinomial logit model, the priors and the design were optimized once 10% of the required sample completed the questionnaire.In study 1, the final experimental design contained 30 choice tasks that were divided into 2 blocks of 15 choice tasks.Each choice task had 2 generic alternatives ("treatment 1" and "treatment 2") that were characterized by a selection of attribute levels, and the third alternative ("no treatment") allowed respondents to not choose any of the alternatives presented (opt-out).The design of the second study consisted of 48 unique choice tasks divided over 3 blocks of 16 choice tasks.Each choice task had 2 generic alternatives.In study 3, the design consisted of 24 choice tasks and was divided into 2 blocks of 12 choice tasks.Again, each choice task had 2 generic alternatives.Examples of the choice tasks are given in Appendix Figures 1 to 3 in Appendix A in Supplemental Materials found at https://doi.org/10.1016/j.jval.2022.05.005.To assess health-risk attitude, we used the 13-item HRAS-13. 16he HRAS-13 is context specific, and its items relate to medical treatment, the importance of health, general attitude toward risk taking in health and care, and consideration of the future consequences of health behaviors.Advantages of using this scale are that (1) context-specific scales are found to better predict risk behavior 17,38 ; (2) it overcomes the discussion about whether risk attitude is context dependent or whether a general risk attitude exists 15,16 ; and (3) it avoids the need to differentiate between risktaking behavior and risk perception. 15,16The items of the HRAS-13 were rated on a 7-point Likert scale from "completely disagree" to "completely agree."Total scores for the HRAS-13 were obtained by reverse-scoring 7 of the items that are phrased negatively and then summing the scores to each item.Scores range between 13 and 91.Respondents with a lower HRAS-13 score are more health risk averse, whereas a higher HRAS-13 score indicates a more riskprone attitude toward health risks.In addition, the questionnaires contained questions about health, age, sex, and education level.Health was measured using a visual analog scale ranging from 0 to 100 in study 1 and study 3, while using a 5-point Likert scale from "very poor" to "very good" in study 2. Age was measured on a continuous scale, with sex as "female," "male," or "other." Education level was measured according to the European Qualification Framework 39 and categorized as "low," "medium," or "high" in accordance with the Dutch Qualification Framework and Statistics Netherlands. 40,41In addition, study 1 and study 2 also contained questions about health literacy and numeracy.Health literacy was measured using the Communicative Health Literacy and Critical Health Literacy scales. 42The scale consists of 5 items on a 4-point Likert scale that ranges from "never" to "always."Based on the average scoring system in de Bekker-Grob et al 43,44 and the 3 categories by Ancillotti et al, 31 respondents with an average score of 2 or lower were categorized as "inadequate," those between 2 and 3 were categorized as "problematic," and those with an average score larger than 3 were deemed "sufficient."The Dutch version of the scale 45 has more items than the original (Japanese) version 42 and the Swedish version. 46In addition, the Dutch version uses a 4-point Likert scale rather than a 5point Likert scale.Study 1 was based on the Dutch version (and translated from Dutch to English and French); study 2 was based on the Swedish version.For comparability between studies, health literacy was calculated using only the 5 items that were in the Swedish version, and responses in study 2 were recoded so that they were also rated on a 4-point Likert scale (divide each item score by 5 and multiply by 4/5).Numeracy was measured using the 3-item version of the subjective numeracy scale. 47Based again on de Bekker-Grob et al 43,44 and Ancillotti et al, 31 items were scored on a 6-point Likert scale ranging from "not good at all/ never" to "extremely good/very often."Respondents with an average score below 2 were categorized as "inadequate," those with a score between 3 and 4 were categorized as "problematic," and those with an average score higher than 5 were deemed "sufficient."

Analysis of Health-Risk Attitude and Preference Heterogeneity
Panel latent class models were used to analyze heterogeneity of preferences.9][50] Following random utility theory, class allocation of respondent n in class c is based on choices for choice set s of each alternative j and is given by U nsjjc .The utility consists of an observable component V and a random component ε nsjjc that is formally written as follows: Here b c is a class-specific vector describing the preference weights of the attributes and levels X 0 nsj for respondent n for choice set s in alternative j.The exact model specification differed per study; the specification of the alternative specific constant(s), linearity of the attributes, and the number of classes were based on model fit and with consideration for class size and interpretability of the maineffects model.To understand whether and how health-risk attitude and preference heterogeneity are related, we analyzed 2 types of heterogeneity, namely, (1) stochastic class assignment and (2) systematic preference heterogeneity.Both types of heterogeneity were included jointly to disentangle the different potential sources of preference heterogeneity.The impact of health-risk attitude on stochastic class assignment was included to analyze whether health-risk attitude could distinguish preferences between classes, that is, whether it distinguished preferences for risk-related attributes and nonrisk-related attributes.For matters of completeness, the class assignment model also included other variables based on their relationship with health-risk attitude or preference heterogeneity.The propensity of class membership f nc is specified as a linear-in-parameters function consisting of a constant term d 0jc plus the variables health 7,16,28 (dichotomized based on median split, good vs rest), age 28,51 (continuous), sex 4,16,28,51 (male vs female), education level 4,28 (high vs rest), and if applicable numeracy 28,51 and health literacy 28,29 (sufficient vs rest); thus: 1jc HRAS score n 1g 2jc good health n 1g 3jc age n 1 g 4jc male n 1g 5jc high education n 1g 6jc sufficient literacy n 1g 7jc sufficient numeracy n 1u nc : (2) The stochastic term w nc is assumed to be extreme value type 1 (Gumbel) independent and identically distributed across classes, yielding a polytomous multinomial logit model for the probability of class membership: Note that the coefficient vector for 1 class must be set to 0. Statistically significant g coefficients (as indicated by P ,.05) indicate that a certain variable contributed to the class assignment model.For example, a positive and statistically significant g coefficient of HRAS score in class 1 would mean that respondents with higher HRAS scores are more likely to be allocated to class 1 than the reference class.Nevertheless, a nonsignificant coefficient means that differences in HRAS scores do not explain differences in overall preference structures between the classes.
In parallel, we assessed the relationship between health-risk attitude and systematic preference heterogeneity by interacting the risk-related attributes with respondents' health-risk attitude.A statistically significant HRAS interaction term (again as indicated by P ,.05) with a risk-related attribute, for example, in class 1, is interpreted as health-risk attitude explaining preference heterogeneity of that attribute within that class.
To assess the impact of including health-risk attitude, in each study, we compared log-likelihood of the model that included health-risk attitude in the class allocation model and used interactions with a model that did not do either but was equal in all other aspects.Log-likelihood statistics were compared using likelihood ratio tests, given that the number of classes is equal between models.All analyses were performed in Nlogit 6.

Respondents
Given the varying study contexts, inclusion criteria, and study designs, the 3 studies had different types of respondents (see Table 1 [30][31][32] for an overview of the case studies).The studies consisted of 753, 378, and 648 respondents, respectively.In study 2, the general public sample, HRAS-13 scores were generally higher (more positive attitude toward health risks) and more dispersed than in the MS sample (study 1) and the HKOA sample (study 3).In study 1, respondents were less healthy (mean = 60.6)than in study 3 (mean = 68.8);they were younger (mean = 42.0),mostly female (67.9%), and highly educated (47.3%).Furthermore, the sample of study 1 was less literate and slightly less numerate than in study 2. In the second study, most people had a good (43.1%) or very good (15.3%)health.The sample of study 2 was slightly older than in the first study, but younger than in the third.As in study 1, most respondents were highly educated (51.3%).In study 3, respondents were oldest (mean = 61.7),55.4% were female, and fewer (25.3%) were highly educated than in the other studies.No data were collected on health literacy and numeracy.An overview of these respondent characteristics can be found in Table 3, whereas more information about the relationship between HRAS-13 scores and other background variables is presented in Appendix Table 1 in Appendix B in Supplemental Materials found at https://doi.org/10.1016/j.jval.2022.05.005.

Health-Risk Attitude and Preference Heterogeneity
An overview of the results per study is presented in Tables 4  and 5 described below; for further information, the full results are presented in Appendix Tables 2 to 4 in Appendix C in Supplemental Materials found at https://doi.org/10.1016/j.jval.2022.05.005.In none of the studies were HRAS-13 scores statistically significantly related to stochastic classification of preferences.This indicates that parameters in the utility function were not jointly dependent on health-risk attitude for any of the classes in any of the studies.Nevertheless, numeracy was related to class allocation (P = .02)in study 1.In study 2, age (P = .004)and health literacy contributed to class allocation (P = .012)in class 1 and 2, respectively.In study 3, age explained class allocation in 2 classes (P = .004and P = .040).
In contrast, systematic heterogeneity as measured by interactions between health-risk attitude and risk attributes was present in some risk attributes of the studies.In study 1, the MS patient sample with 3 risk attributes phrased using percent, we found systematic preference heterogeneity for all risk attributes in the first and largest class.In this class, health-risk attitude significantly moderated the effect of reducing the risk of relapse and reducing disease progression (P , .001 for both) and the risk of rare severe side effects (P = .020).In the second class, only the interaction between health-risk attitude and reducing risk of relapse (P = .003)was significant.In addition, the second study, concerning the antibiotics context with a general public sample, had 3 classes and 3 risk attributes.Health-risk attitude explained part of the heterogeneity for treatment failure rate (P = .019)in one of the classes but not in the other 2. Nevertheless, the interaction effects with the other risk attributes were not significant in any of the classes.In the third study about patients' preferences for HKOA treatment, 2 attributes implicitly concerned risk.In 2 of the 4 classes, health-risk attitude explained heterogeneity of preferences for waiting time (P = .001for both) but not for professionals involved.
As shown in Table 5, inclusion of HRAS-13 scores significantly improved the model fit only in study 3 (c 2 = 37.9, df = 15, P , .001).In the other studies, the improvement was not statistically significant.

Discussion and Conclusions
Hence, where does health-risk attitude come into the equation when researching preference variation?Our study did not find evidence that health-risk attitude as measured by the HRAS-13 distinguishes people between classes.Nevertheless, we did find evidence that the HRAS-13 can distinguish people's preferences for some risk attributes within classes.This association between health-risk attitude and preference heterogeneity was stronger in the case studies where respondents were sampled from a patient population than in the case study that used a general public sample.Respondents in the patient samples were also more health risk averse than members of the general public.In the first case study, which used a patient sample, health-risk attitude explained the heterogeneity of preferences for most attributes in both classes, but it did not significantly improve the model fit.In the third study, which also used a patient sample, health-risk attitude was related to heterogeneity of preferences for one attribute in 2 of 4 health preference classes.Although the 2 risk attributes of this study only implicitly concerned risk, it was the only study in which the model fit statistically significantly improved by incorporating health-risk attitude.
Furthermore, we found that numeracy, health literacy, and age affected stochastic class allocation, meaning that these characteristics could distinguish preferences between classes for riskrelated attributes and nonrisk-related attributes.In the study where numeracy affected class allocation, all risk attributes were phrased using percent.In the study where health literacy affected class allocation, one of the risk attributes was described in words.Moreover, numeracy and literacy were among the characteristics that improved external validity when accounted for in preference heterogeneity in de Bekker-Grob et al 28 and among the psychological constructs with the strongest consensus to be included in preference studies in the review of Russo et al. 12 Our results suggest that risks are in some way related to preference heterogeneity, either directly when health-risk attitude distinguishes people's preferences within classes or indirectly when people have varying levels of numeracy and literacy.
The relevance of these results is threefold.First, the impact of health-risk attitude on preferences should be explored beyond class membership by interacting the health-risk attitude with the risk-related attributes.This is expected to be mostly relevant in contexts where alternatives largely vary in terms of benefit-risk, when treatment outcomes are highly uncertain, or when patients are risk averse.In those contexts, accounting for health-risk attitude has the potential to improve model fit and model interpretations.Second, the impact of health-risk attitude on preferences should be explored using a different measure than the HRAS-13.Given that we did not find strong evidence for this using the HRAS-13, which is a health-specific instrument of which the items cover a broad range of health domains, 16 an option would be to use a more targeted measure of health-risk attitude in DCEs.In addition, one could research the relationship between health preference heterogeneity and measures that use a narrower definition of risk attitude (eg, the standard gamble method [52][53][54] or the Balloon Analog Risk Task 55 ).Such studies can confirm whether indeed health-risk attitude is not linked to preferences as strongly as anticipated 12,13 or whether it could be explained by the relatively low levels of variance in the HARS-13 scores in the case studies.As outlined in the Methods section, we do recommend sticking to a health-specific measure of risk attitude.Third, given that numeracy and health literacy were found to affect stochastic class allocation, our results add to existing literature that stresses the importance of the communication of risks (ie, presentation, framing, training materials, and analysis) in DCEs (eg, Harrison et al, 22 Veldwijk et al, 29 and Peters et al 56 ).In this study, we analyzed a wide range of risk attributes.Although we did not observe clear differences in the relationship between health-risk attitude and preference heterogeneity based on the type or phrasing of the risk attribute, we find that numeracy explained heterogeneity in the study in which risks were presented using percent, whereas literacy explained heterogeneity in the study where some risk attributes were phrased using percent and some using words.
A strength of this study is that it is among the first to research health-risk attitude as an individual characteristic underlying heterogeneity in health preferences and thereby responds to the call for this type of research. 12,13,22The case studies provide a cross-European comparison in 3 different health contexts with varying degrees of risk and study population leading to an increased generalizability of the results.Nevertheless, the differences in samples also make it harder to identify the source of similarities and differences in results between the studies.Given that secondary data were used for the current study, comparability across the studies is limited.In future research, it would be interesting to set up studies with the specific aim to compare the impact of health-risk attitude across different populations and risk attributes.It should also be noted that it is unclear whether respondents' level of perceived riskiness of the attributes is in line with what was determined by the researchers.Given that risk perception and risk behavior are not always aligned, 15 we recommend future research in this area to also elicit respondents' risk perception at an early stage of DCE development.Furthermore, this research focused on improving model fit and model interpretations from the perspective of internal validity.Given the mixed evidence regarding the predictive ability of survey-based measures of risk attitude, 17,18,38 it would be interesting to also study whether and how health-risk attitude and heterogeneity of health preferences are related from the perspective of external validity and individual-level prediction accuracy, for example, as in de Bekker-Grob et al. 43 In conclusion, our study did not find evidence that health-risk attitude as measured by the HRAS-13 distinguishes people -between classes.Nevertheless, we did find evidence that the HRAS-13 can distinguish people's preferences for risk attributes within classes.This phenomenon was more pronounced in the patient samples than in the general population sample.Furthermore, we found that preference heterogeneity is affected by numeracy and health literacy.These results warrant the relevance of further research into preference heterogeneity beyond class membership, a different measure of health-risk attitude, and the communication of risks.

Table 1 .
Case study and DCE design characteristics.Note.Attributes and levels were selected based on literature review, focus groups, and interviews; they are presented in Table2.
DCE indicates discrete choice experiment; HKOA, hip and knee osteoarthritis; MS, multiple sclerosis.

Table 2 .
Attributes and levels.
GP indicates general practitioner; HKOA, hip and knee osteoarthritis; kr., Swedish Krona; MS, multiple sclerosis.*Attributes with an asterisk implicitly or explicitly concerned risks.

Table 3 .
Respondent characteristics per study.

Table 4 .
Overview results per study.

Table 5 .
Model fit per study.HKOA indicates hip and knee osteoarthritis; HRAS, Health-Risk Attitude Scale; MS, multiple sclerosis.