1. Introduction
Given their knowledge of healthcare and available resources, patients, clinicians, payers, the industry, and policymakers all seek to create the best structures, processes, and outcomes (
Donabedian 1988). Each decision maker aims at maximizing the benefits while minimizing the risks and cost. Healthcare systems are facing a paradigm shift towards value-based healthcare (
Porter and Teisberg 2006). Health preference research (HPR) is the subfield of health economics dedicated to understanding the value of health and health-related objects using observational or experimental methods. Understanding the value of health and healthcare can inform decisions at many different levels of healthcare decision making. These include basic drug discovery and clinical trials, patient-level care, and policymaking for the community by a range of stakeholders and decision makers (
Rubio et al. 2010).
By definition, the theoretical framework underlying HPR rests on two constructs: “health” and “preference.” Health, according to the
World Health Organization (
1946), is “a state of complete physical, mental and social well-being and not merely the absence of disease or infirmity”. Like any construct, one person’s definition of “health” may differ from another’s, and as such, it is always measured with a degree of error; describing it also requires a broad array of indicators. Preference requires a comparison and may be defined as “a greater liking for one alternative over another” (
Simpson and Weiner 1989). Health preference researchers typically assume that people prefer better and/or more health to worse and/or less health, and behave accordingly when they make health-related decisions. To understand the value of health and healthcare, a researcher examines choices by patients, clinicians, payers, policymakers, or anyone engaged in health-related decisions.
In HPR, health and preference are linked together through choices—the unit of observation in a health preference study. Formally, choice refers the selection of an object from among two or more alternatives. When we observe an individual’s choices, this behavior implicitly reveals their ordering of alternatives or objects (
White 1969). Researchers typically assume that each choice between health-related alternatives maximizes decisional utility (
Luce 1959;
Marschak 1960). Health preference studies use observational and experimental methods to collect empirical evidence on health-related choices.
A discrete choice experiment (DCE) is a behavioral study where respondents are allocated a series of choice sets under a hypothetical situation. In each task, the respondent is asked to choose among two or more alternatives in each choice set. As part of the DCE protocol, their choices are observed, captured, recorded, and analyzed as a means to test one or more hypothesized causal relationships. Often, alternatives are described using a descriptive system of multilevel attributes, and the choice sets may include a referent alternative (e.g., opt-out). In health economics, the primary purpose of a DCE is to test and estimate the causal relationships between attribute levels and preferential choice behaviors. However, in DCEs conducted in other fields, such as audiology, the primary purpose of DCE is to assess the effects of a stimulus on perceptions. Apart from causal relationships, health economists conduct DCE to examine the willingness to pay for alternatives or predict choice behaviors under hypothetical situations (e.g., health technology assessment).
Between DCE respondents, the effects of attributes on behavior may vary (i.e., interpersonal heterogeneity) due to differences in individual utilities. Such differences in utilities between persons may be stochastic or systematically associated with observable or unobservable factors. Observable heterogeneity refers to observed (by the researcher) factors (e.g., age) that are associated with observed effects (also known as explained heterogeneity). On the other hand, unobservable heterogeneity refers to latent factors associated with the unobserved effects (e.g., risk perception). Multiple researchers stated that a failure to account for heterogeneity may lead to egregious differences between the estimated and true values in a choice analysis (
Bhat 1998;
Brownstone and Train 1999;
Chamberlain 1980;
McFadden and Train 2000). However, the extent of such differences is unclear.
Like in panel data analysis more generally, there are two general approaches to account for interpersonal heterogeneity depending on the experimental design. In some cases, the DCE has a large number of tasks per respondent relative to the number of parameters; therefore, the analyst can estimate individual-specific parameters per respondent (namely, fixed effects). When this is not the case, interpersonal heterogeneity is unobservable; nevertheless, the analyst may estimate a random parameter model by making distributional assumptions about the prevalence of parameters within the population (e.g., mixed logit). In this simulation study, we examine the implications of ignoring interpersonal heterogeneity. We simulated a standard DCE under various forms of unobservable interpersonal heterogeneity and assessed the implications of ignoring this heterogeneity. More specifically, we generated trinomial responses for a standard DCE with a full factorial design and individual-specific coefficients varying over individuals. However, we ignored heterogeneity and estimated the conditional logit (CL) model to examine the differences between the estimated and true coefficients.
This paper shows what happens when different types of heterogeneity are ignored, and the implications for health economics, where conditional logit remains to be a standard modeling approach. By comparing the results with the true values, this paper clarifies whether ignoring certain types of heterogeneity alters the interpretation of willingness-to-pay (WTP) and choice predictions. For simplicity, the results of this study are limited to just one DCE and may lead other researchers to identify similar patterns under different experimental designs or types of heterogeneity (e.g., latent classes).
1.1. Types of Unobservable Heterogeneity
In HPR, unobservable interpersonal heterogeneity may be classified as either informative or uninformative. Uninformative heterogeneity implies that parameters differ between persons, but knowing one parameter of a person does not inform about the person’s distribution of another parameter (i.e., uncorrelated parameters). Under this form of uninformative heterogeneity, individual-specific parameters vary between persons, but these parameters are not correlated. Conversely, informative heterogeneity implies that the individual-specific parameters are correlated through latent factors.
Furthermore, informative heterogeneity may be separated into two subtypes on the basis of which parameters are correlated. Substitution patterns (e.g.,
Brownstone and Train 1999) refer to the presence of a latent factor that causes an association with alternative-specific constants (ASCs). For example, the classical red-bus–blue-bus problem refers to a positive correlation between the two bus alternatives, where a red bus is a substitute for a blue one (e.g.,
Quandt 1970). An ASC is not the same as a nominal attribute. In this classical problem, every set includes three fixed alternatives, namely, a car, a red bus, and a blue bus. If the color and mode were nominal attributes and varied independently, some sets might include three buses. In health economics, a more common example is a choice set with multiple fixed treatment options (e.g., a brand-name medication and a generic medication) and “watchful waiting” (i.e., no medication).
Taste patterns (e.g.,
Revelt and Train 1998) refer to the presence of a latent factor that causes an association between attribute coefficients more generally (i.e., attribute importance). In health economics, each alternative may vary by out-of-pocket price, wait times, and other attributes. The coefficients of these attributes may be correlated. For example, individuals who are willing to pay more for their healthcare may also prefer more luxurious services (e.g., a brand-name medication when a generic medication is available).
Imagine a specific alternative (e.g., a brand-name medication). Under uninformative interpersonal heterogeneity, the effect of a specific brand varies between persons, but this effect is not associated with another alternative’s effect (i.e., substitution pattern) or the effects of any attributes (i.e., taste pattern). Alternatively, latent factors may induce informative heterogeneity, such as a substitution pattern between brands or a taste pattern between a brand and out-of-pocket cost.
For simplicity, we generated preferential choice behaviors for a standard DCE with a full factorial design (
Appendix A) and individual-specific unobserved coefficients drawn from a normal distribution. Each task represents a choice among three alternatives to illustrate uninformative and informative heterogeneity, including substitution and taste patterns. Each alternative is described using a nominal and a numeric attribute, namely, medication and out-of-pocket price. For example, the correlation between the effects of brand and generic is a substitution pattern, and the correlation between the effects of medication and price is a taste pattern. Willingness to pay for a brand-name medication is the ratio of the brand-name medication and price effects. This DCE was constructed to mimic the classic red-bus–blue-bus example.
1.2. Choice Modeling with Unobservable Heterogeneity
Historically, econometricians estimated the parameters of CL models under the assumption of homogeneity. When heterogeneity is absent, the identification of the CL is standard and its nuances are beyond the scope of this paper. Choice probabilities of the CL model have a closed-form which can be evaluated analytically using the maximum likelihood (ML) estimator. When heterogeneity is present and observable, the CL model can be adapted to “explain” this heterogeneity, and its ML estimation is again computationally simple. However, one of the drawbacks of the CL model is that it does not account for unobservable heterogeneity. Failing to account for unobservable heterogeneity (i.e., uninformative and informative) may lead to misleading statements about the WTP and choice predictions (
Bhat 1998;
Brownstone and Train 1999;
Chamberlain 1980;
McFadden and Train 2000). There are more flexible models that allow for unobservable heterogeneity, such as mixed logit specifications; however, our main purpose is to understand consequences of ignoring unobservable heterogeneity.
For this simulation study, we explore the differences between CL estimates and the mean of their true values depending on the type of unobservable heterogeneity. For the purposes of this paper, we define bias as the difference between the estimates of the CL under ML and the mean of the true values used to generate the responses. First, we examine the case of omitted variance bias where one or more parameters vary, but are uncorrelated. Understanding such uninformative heterogeneity is a necessary precursor to the examination of informative heterogeneity (i.e., correlations). Second, we examine the three cases of omitted correlation bias: substitution patterns, taste patterns, and both.
There are more flexible models of DCE responses, such as mixed logit specifications, which allow for unobservable heterogeneity. However, our main purpose is to understand consequences of ignoring unobservable heterogeneity. Furthermore,
Clark et al. (
2014) reported that, in health economics, 44% of DCE studies during 2009–2012 used the CL (or multinomial logit) model, while 21% used the mixed logit model during the same period. Our findings have intriguing implications for the majority of health preference studies, including benefit–risk analyses and health valuations.
The rest of the paper is organized as follows.
Section 2 presents the CL model, and the simulation and estimation details.
Section 3 presents the results and
Section 5 concludes.
2. Methods
2.1. Model
The theoretical framework of this choice analysis is based on random utility maximization (RUM) theory (
Marschak 1960). According to RUM theory, utility function
of individual
for alternative
at choice situation
can be decomposed into a deterministic part of utility
(representative utility) which is observed by the researcher, and a random part of utility
that is not observed by the researcher. Following
McFadden (
1974), individual
i chooses an alternative
j if and only if the probability that utility associated with alternative
j is higher than that of utilities associated with all other alternatives:
Choice probabilities are calculated on the basis of a relative measure where utility of one of the alternatives in the choice set is taken as a reference. To derive the choice probabilities, distributional assumptions are needed about the random part of utility. The CL model is derived under the assumption that
is independently and identically distributed (IID) with extreme value type I (EV1) distribution (e.g.,
Brownstone and Train 1999;
McFadden and Train 2000;
Revelt and Train 1998;
Train 2009). As a result, the difference between two IID EV1 random error terms (
) has logistic distribution. This implies that the choice probabilities of the CL model can be expressed in terms of logistic distribution with cumulative distribution function:
In this simulation study, we generated choices among three alternatives, where the third alternative served as a referent (i.e., opt-out) (e.g.,
Campbell and Erdem 2019). For example, imagine the decision on whether to take a heartburn medication: brand-name, generic, or no medication. Hence, we assumed that the representative utilities of the three alternatives took the following form:
Therefore, utility differences take the following form:
where
and
are out-of-pocket prices (see
Appendix A for the out-of-pocket price),
and
are alternative-specific constants (ASCs) for each respondent,
is an individual-specific price coefficient,
and
are two utility differences, and finally,
and
are differences in IID EV1 error terms. Under this specification, the willingness to pay for each medication is the ratio of the ASC and price coefficient,
.
In our model specification, we used two ASCs ( and ) to allow their correlation. Alternatively, a nominal attribute for opt-in () may be included instead of two ASCs. However, that specification would deviate from the red-bus–blue-bus example. Similar in health economics is the case of choosing among three ways to see heartburn relief, where the alternatives are a brand-name medication, a generic medication, or doing nothing. It seems appropriate to allow the ASCs for the brand-name and generic medications to be correlated and not combined as a treatment attribute.
To incorporate interpersonal heterogeneity, we express the prevalence of individual-specific ASCs and
within the population using a multivariate normal distribution with
. Covariance matrix
is
where
,
,
,
,
and
. Uninformative heterogeneity occurs when correlation parameters are restricted to be equal to zero
, in which case, covariance matrix
is diagonal. On the other hand, informative heterogeneity allows for the correlation of individual-specific parameters, such that
.
2.2. Simulation
In the data generation process, the true means of the representative utility parameters (Equation (
3)) were fixed at
and
(i.e., brand-name, generic, and out-of-pocket price, respectively). The out-of-pocket price values that we used in this study (
and
) are presented in
Appendix A. Furthermore, we generated various covariance matrices (Equation (
5)) on the basis of the true values that we chose for the variances and correlations, and we refer to each distinct covariance matrix as a “parameter set” (see
Table 1).
The simulated covariance matrices fall into six groups: the first three groups represent uninformative heterogeneity (i.e., uncorrelated parameters;
), and the last three groups represent informative heterogeneity (correlated parameters;
) where the variances were fixed at 0.15 for identification. For instance, in Group 2, we assumed that
and
ranged from 0.025 to 0.25 with increments of 0.025 (i.e.,
or
), and we assumed that
. Since there were 10 different values for each
and
, we could create 100 combinations of parameter sets for the true covariance matrix. The rest of
Table 1 can be interpreted in a similar fashion.
Across all simulations, the ratio of the mean (i.e., ) brand-name medication and price parameters was −1. Under homogeneity, this ratio represents WTP. Under heterogeneity, WTP is the mean of the true ratios (i.e., ), and its value depends on the correlation . In Group 4, the true mean WTP for the brand-name medication (A) monotonically decreased from −1.011 to −1.337 as the correlation increased from −0.95 to 0.95. This association is intuitive: a large negative correlation implies that the variations in and counterbalance each other (i.e., when is less than −1, and is more likely to be greater than 1). Likewise, a positive correlation implies that larger are more likely divided by a smaller , causing the mean WTP to be lower (i.e., <−1). Even in the absence of correlation, the mean WTP was around −1.2. By construction, CL is unable to estimate both mean parameters and willingness-to-pay consistently under interpersonal heterogeneity; however, the extent of the bias is unclear.
At first glance, Group 6 seems to contain 4000 combinations of parameters, since there are 10 different values of , and 20 different values of and (10 × 20 × 20); however, we eliminated parametric combinations that did not lead to a positive definite covariance matrix , reducing the number of sets to 2520.
For each parameter set, we simulated 1000 DCE datasets (
), each including 25 choices (
) for 200 respondents (
) given a full factorial experimental design (see
Appendix A). This design may include dominated alternatives. However, since this is a simulation study, the existence of dominated alternatives should not be a concern. With each simulated dataset, we estimated a CL model by ML, and compared the estimated and mean true parameters (
,
and
). In addition, we calculated the two WTP on the basis of the ratio of estimated parameters and predicted the 75 choice probabilities in the experimental design (3 alternatives by 25 choice set). We compared the true and estimated choice predictions (
P) across the 1000 simulated datasets using Lin’s concordance correlation coefficient, which we denote as Lin’s
(
Supplementary Materials). Furthermore, we calculated the worst-case scenario in the results across all simulations on the basis of the ratio of estimated and true odds (
Supplementary Materials).
4. Discussion
These simulation results exemplify key implications of ignoring interpersonal heterogeneity in DCE, namely, the omitted variance and correlation biases. The results support the belief that assuming homogeneity in the presence of heterogeneity produces inconsistent results (see
Chamberlain 1980). Specifically, the CL parameter estimates (
,
and
) differed from their true values under certain specifications of the data generation process. Likewise, we found substantive differences between the CL and true mean WTP. For instance,
Table 3 shows a 20% reduction in WTP (i.e., [1.337 − 1.069]/1.337 = 0.2). Interpersonal heterogeneity generally implies that the CL enlarges the parameter estimates and increases the WTP estimates (i.e., less negative). Nevertheless, these differences largely factor out in the choice probabilities.
Within health economics, the findings support the conclusion that CL WTP estimates for medical treatments may be underestimated due to interpersonal heterogeneity. Furthermore, the assumption of preference homogeneity among patients (i.e., “misalignment with a patient’s wishes”) may contribute to an underestimation of the value of treatment, substantially reducing quality of care (
Clapp et al. 2022). If the analyst conducted the DCE to solely predict choices in aggregate, the presence of interpersonal heterogeneity may not be as relevant. CL choice probabilities were similar to the true ones. The reason why the biases in the estimated coefficients did not transfer to the choice probabilities may have been because the influences of the interpersonal heterogeneity on the likelihood of choices reduce significantly when we pass the effects through the logit cumulative distribution function.
Although we characterized the estimated ratio as WTP (i.e., the ratio of the mean brand-name medication and price parameters), many health preference studies estimate similar ratios for other purposes. In a benefit–cost analysis, results are often expressed as the maximal acceptable risk (i.e., the ratio of the mean benefit and risk parameters) (
Reed et al. 2022). In health valuation, values are expressed on a quality-adjusted life year scale (e.g., the ratio of a mean EQ-5D episode and one year in full health) and used in economic evaluations to inform resource allocation decisions (
Craig and Rand 2018). Like with WTP, the maximal acceptable risk and EQ-5D value sets may be biased due to interpersonal heterogeneity among the respondents.
Previous research on interpersonal heterogeneity showed that the mixed logit model may perform better than the CL model in terms of model fit (e.g.,
Bhat 1998;
Brownstone and Train 1999;
McFadden and Train 2000). A natural extension to this paper would be to estimate the mixed logit using the simulated DCE and demonstrate whether it mitigates the biases in the parameters and WTP estimate shown. Furthermore, we may examine how well the mixed logit performs in the estimation of the variance and correlations under alternative specifications. A future study may also examine subsets of the full factorial design to assess the implication of fractional factorial designs or unbalanced designs (i.e., where set-specific sample sizes vary) on the estimation of the mixed logit. Overall, further research is needed to fully grasp the circumstances that interpersonal heterogeneity affects DCE interpretations.
We recognize three primary limitations of this simulation study. First, the results are based on a single DCE with 3 alternatives, 3 attributes, 200 responses per choice set, and a full factorial design. This DCE was simulated for its simplicity and capacity to mimic multiple forms of unobservable heterogeneity. Future work may attempt to evaluate these findings using other DCE simulations. Second, the results are based on normally distributed random coefficients. Alternatively, the random coefficients may be distributed asymmetrically or discretely (i.e., latent classes). Lastly, a future study may wish to account for both interpersonal heterogeneity and intrapersonal variability (i.e., how a person’s parameters vary between choices). Each respondent’s parameters may be random with its own mean and variance-covariance matrix, creating a more flexible function form (e.g., hierarchical Bayes).