A Comparison of Different Bayesian Design Criteria for Constructing Efficient Discrete Choice Experiments
Speaker
Abstract
| Bayesian optimal design theory provides a solid basis for coping with the problem of design dependence on the unknown parameters in stated preference studies. The Bayesian design criterion that is used in published work on the optimal design of stated preference studies is based on the Fisher information matrix. However, several other Bayesian design criteria exist, some of which are known to have better finite sample properties than criteria based on the Fisher information matrix. The alternative design criteria are based on the generalized Fisher information matrix, the expected posterior covariance matrix, and the expected gain in Shannon information. In this study, we apply these alternative Bayesian design criteria in the context of stated preference studies and compare the performance of the resulting stated preference designs. We investigate in detail how well the designs perform in terms of the design criteria for which they were not optimized, and study situations where the stated preference data are analyzed in a Bayesian fashion and in a non-Bayesian fashion (using maximum likelihood). Our simulation results favor a Bayesian design criterion based on the generalized Fisher information matrix, as it appears to be the only computationally feasible criterion that can compete with the overall best criterion, which is based on the expected posterior covariance matrix. |
| Contact information: |
| Erik Kole |
