Some thoughts on sample size: a Bayesian-frequentist hybrid approach.

BACKGROUND: Traditional calculations of sample size do not formally incorporate uncertainty about the likely effect size. Use of a normal prior to express that uncertainty, as recently recommended, can lead to power that does not approach 1 as the sample size approaches infinity.

PURPOSE: To provide approaches for calculating sample size and power that formally incorporate uncertainty about effect size. The relevant formulas should ensure that power approaches one as sample size increases indefinitely and should be easy to calculate.

METHODS: We examine normal, truncated normal, and gamma priors for effect size computationally and demonstrate analytically an approach to approximating the power for a truncated normal prior. We also propose a simple compromise method that requires a moderately larger sample size than the one derived from the fixed effect method.

RESULTS: Use of a realistic prior distribution instead of a fixed treatment effect is likely to increase the sample size required for a Phase 3 trial. The standard fixed effect method for moving from estimates of effect size obtained in a Phase 2 trial to the sample size of a Phase 3 trial ignores the variability inherent in the estimate from Phase 2. Truncated normal priors appear to require unrealistically large sample sizes while gamma priors appear to place too much probability on large effect sizes and therefore produce unrealistically high power.

LIMITATIONS: The article deals with a few examples and a limited range of parameters. It does not deal explicitly with binary or time-to-failure data.

CONCLUSIONS: Use of the standard fixed approach to sample size calculation often yields a sample size leading to lower power than desired. Other natural parametric priors lead either to unacceptably large sample sizes or to unrealistically high power. We recommend an approach that is a compromise between assuming a fixed effect size and assigning a normal prior to the effect size.