OBJECTIVES: To compare how the Dirichlet and standard discrete choice (multinomial logit, MNL) model differ in their convergence to stable estimates of population benefit-risk preferences.

METHODS: Individual-level trade-offs from a three-attribute patient preference study (n=560) were used in a simulation study where MNL model was applied to individual choice data, and Dirichlet distribution applied to individual trade-off data. Discrete choice experiments (DCE) were conducted for sample sizes from 20 to 540 (100 simulations per sample size), where in each one subjects drawn from the empirical distribution were simulated to answer 6 DCE questions. MNL models were fit to the answer data and Dirichlet distribution estimated for the trade-off weights. Convergence was measured with Euclidean distance of the obtained estimate and the estimate with the full sample. Logit model variability was additionally assessed with coefficient p-values.

RESULTS: The full-sample mean estimate of the Dirichlet model ([0.52, 0.17, 0.31]) was close to the experimental mean ([0.54, 0.14, 0.32]), whereas MNL was considerably different (normalized mean [0.57, 0.06, 0.36]). Both methods converge towards the full-sample estimate, and have <0.05 distance from the full-population estimate with a sample size of at least 260 (MNL) and 160 (Dirichlet) in 95% of the simulations. With MNL, in many simulations the preference estimate of the least important attribute is still insignificant (p>0.05). For example, with a sample size of 200, in 9% of the simulations, the analyst would not be able to conclude significance of the estimate.

CONCLUSIONS: Dirichlet distribution is likely to have smaller sample size requirements than the MNL model, and is a promising approach for modeling population preferences.

Disclaimer: The views presented here are those of the authors and not of their organizations.