From Individual to Population Preferences: Comparison of Discrete Choice and Dirichlet Models for Treatment Benefit-Risk Tradeoffs.

Introduction. The Dirichlet distribution has been proposed for representing preference heterogeneity, but there is limited evidence on its suitability for modeling population preferences on treatment benefits and risks. Methods. We conducted a simulation study to compare how the Dirichlet and standard discrete choice models (multinomial logit [MNL] and mixed logit [MXL]) differ in their convergence to stable estimates of population benefit-risk preferences. The source data consisted of individual-level tradeoffs from an existing 3-attribute patient preference study (N = 560). The Dirichlet population model was fit directly to the attribute weights in the source data. The MNL and MXL population models were fit to the outcomes of a simulated discrete choice experiment in the same sample of 560 patients. Convergence to the parameter values of the Dirichlet and MNL population models was assessed with sample sizes ranging from 20 to 500 (100 simulations per sample size). Model variability was also assessed with coefficient P values. Results. Population preference estimates of all models were very close to the sample mean, and the MNL and MXL models had good fit (McFadden's adjusted R2 = 0.12 and 0.13). The Dirichlet model converged reliably to within 0.05 distance of the population preference estimates with a sample size of 100, where the MNL model required a sample size of 240 for this. The MNL model produced consistently significant coefficient estimates with sample sizes of 100 and higher. Conclusion. The Dirichlet model is likely to have smaller sample size requirements than standard discrete choice models in modeling population preferences for treatment benefit-risk tradeoffs and is a useful addition to health preference analyst's toolbox.