may not be able to complete the tasks. Where there is suspicion that disability may affect people then it makes sense to simplify the conjoint tasks as much as is reasonable.

It is important to describe the subject-specific health and demographic characteristics of the sample and compare these characteristics to the population to which you wish to generalise. Reviewers and readers inevitably will question whether you have inadvertently captured the views of more highly educated, more proactive patients (such as those who participate in patient advocacy groups) or patients with higher-than-average incomes.

viii. Statistical analyses

Conjoint analysis data and the modeling of preferences can require some complex statistical methods, hence it is vital to ask, “Were statistical analyses and model estimation conducted appropriately?”

There are several objectives of analyzing stated-preference data. First, one wants to estimate the strength of preference for the attributes and attribute levels included in the survey. One might also be interested in estimating how preferences vary by individual-subject characteristics. For policy analysis, one might also be interested in calculating how choice probabilities vary with changes in attributes or attribute levels, or in calculating secondary estimates of money equivalence (WTP) (42), risk equivalence (maximum acceptable risk (MAR)) (43), or time equivalence for various changes in attributes or attribute levels (44).

Theoretically valid and unbiased preference estimates depend on model specifications that are consistent with the underlying utility theory used to elicit preferences and with the particular features of the response and profile variables. Forced-choice conjoint analyses, discrete-choice experiments, and rating studies lend themselves to analysis using stochastic utility maximization theory. Rating or card-sort conjoint data often are analyzed using ordinary least squares or ordered-probit methods.

In many conjoint analyses, multiple responses are obtained from each subject. In these cases, researchers should ensure that the statistical analysis of the conjoint data account for within-subject correlation. Ignoring the fact that each subject provides multiple responses can result in biased preference estimates. Thus, researchers who estimate these models should test that the data being analyzed is consistent with the assumptions required for the model being employed.

Researchers must determine whether to model attribute levels as continuous or categorical. If attribute levels are specified as continuous, researchers must determine the appropriate functional form for each continuous variable. Categorical models avoid imposing any functional form on preference weights and provide a validity check on the correct ordering of naturally ordered attribute levels. In addition, researchers should determine whether categorical attribute levels are specified as dummy variables or effects-coded variables. When effects coding is used, zero corresponds to the mean effect for each attribute, rather than the combination of all the omitted categories, and the parameter for the omitted category is the negative sum of the included-category parameters. Hensher et al. (13) explain why effects coding is statistically superior for choice models.

Finally, researchers also should account for differences in preferences that arise from differences in individual characteristics such as age, income, education, and gender by interacting individual characteristics with attributes

# ISPOR Conjoint Analysis in Health Task Force Report

Page 12