Blending Bayesian and frequentist methods according to the precision of prior information with applications to hypothesis testing

The proposed minimax procedure blends strict Bayesian methods with p values and confidence intervals or with default-prior methods. Two applications to hypothesis testing bring some implications to light. First, the blended probability that a point null hypothesis is true is equal to the p value or a lower bound of an unknown posterior probability, whichever is greater. As a result, the p value is reported instead of any posterior probability in the case of complete prior ignorance but is ignored in the case of a fully known prior. In the case of partial knowledge about the prior, the possible posterior probability that is closest to the p value is used for inference. The second application provides guidance on the choice of methods used for small numbers of tests as opposed to those appropriate for large numbers. Whereas statisticians tend to prefer a multiple comparison procedure that adjusts each p value for small numbers of tests, large numbers instead lead many to estimate the local false discovery rate (LFDR), a posterior probability of hypothesis truth. Each blended probability reduces to the LFDR estimate if it can be estimated with sufficient accuracy or to the adjusted p value otherwise. Copyright Springer-Verlag Berlin Heidelberg 2015