Design Considerations for Bayesian Clinical Studies: Prior Effective Sample Size and Type 1 Error Level
*Gene Anthony Pennello, FDA/CDRH  Laura Thompson, FDA 

Keywords: hierarchical model, adaptive design, interim look,

Bayesian studies with informative priors require special consideration at the design stage. For example, how influential is the prior distribution? How is sample size determined? What is an acceptable level of type 1 error? One measure of prior influence is the prior effective sample size (PESS), the effective number of subjects borrowed from the prior distribution. If PESS is defined based on comparing the prior with the likelihood on Fisher information, reasonable expressions are obtained for conjugate priors and hierarchical models. Moreover, Bayesian sample size can be determined to be approximately the frequentist sample size minus PESS, where frequentist sample size is evaluated at the prior mean or mode. The extra sample size contained in the prior suggests that the type 1 error rate for a Bayesian study should be calculated “prior to the prior”. If type 1 error rate is based as usual on the study data alone, then an analysis controlling this rate will nullify any advantage gained in utilizing the prior information. Regarding the prior as extra sample size suggests new directions in Bayesian design, e.g., an interim look can be taken not at a pre-specified samples size but at a pre-specified effective sample size (ESS) of the posterior distribution, which is just the PESS for future data.