While Bayesian G- and V-optimal designs for the multinomial logit model have been shown to have better predictive performance than Bayesian D- and A-optimal designs, the algorithms for generating them have been too slow for commercial use. In this article, we present a much faster algorithm for generating Bayesian optimal designs for all four criterial while simultaneously improving the statistical efficiency of the designs. We also show how to augment a choice design allowing for correlated parameter estimates using a sports club membership study.
- Alternating sample algorithm
- Bayesian D-, A-, G-, and V-optimality
- Conjoint choice design
- Coordinate-exchange algorithm
- Minimum potential design
- Multinomial logit