Bayesian D-Optimal Design Issues for Binomial Generalized Linear Model Screening Designs

Bayesian D-optimal designs have become computationally feasible to construct for simple prior distributions. Some parameter values give rise to models that have little utility to the practitioner for effect screening. For some generalized linear models such as the binomial, inclusion of such models can cause the optimal design to spread out toward the boundary of the design space. This can reduce the D-efficiency of the design over much of the parameter space and result in the Bayesian D-optimal criterion's divergence from the concerns of a practitioner designing a screening experiment.