2023 年 44 巻 1 号 p. 35-51
Determination of the number of subjects to include in a clinical trial is a crucial aspect of experimental design. The standard methodology for sample size determination (SSD) has been established based on a frequentist perspective, while the literature addressing the SSD problem from a Bayesian perspective has increased for the last 20 years. In this paper I discuss the basic concept of Bayesian SSD, with specific focus on an inferential performance-based (non-decision theoretic) approach, using two distinct prior distributions: analysis prior and design prior. The analysis prior formalizes pre-trial information, and it is used to obtain posterior distributions, while the design prior describes a scenario and it is used to obtain prior predictive distributions. In practice, the specification of prior distributions is a key element of Bayesian inference. The prior information may be derived from either expert beliefs or relevant empirical data, and the subjective knowledge derived from an expert elicitation procedure may be useful to define a prior distribution when no or limited data from previous studies is available. In experimental design, the interplay between Bayesian and frequentist methodology is intrinsic. Whichever method is used in SSD, the distinction between demands as expressed in the range of equivalence, and their expectation or beliefs, as represented by the prior information is of paramount importance.