A Statistical Framework for Replicability
We introduce a novel statistical framework to study replicability which simultaneously offers overall Type-I error control, an assessment of compatibility and a combined confidence region. The approach is based on a recently proposed reverse-Bayes method for the analysis of replication success. We show how the method can be recalibrated to obtain a family of combination tests for two studies with exact overall Type-I error control. The approach avoids the double dichotomization for significance of the two-trials rule and has larger project power to detect existing effects. It gives rise to a p-value function which can be used to compute a confidence region for the underlying true effect. If the effect estimates are compatible, the resulting confidence interval is similar to the meta-analytic one, but in the presence of conflict, the confidence region splits into two disjoint intervals. The proposed approach is applied to data from the Experimental Economics Replication Project.
READ FULL TEXT