Directional Adaptive Filtering for Replicability Analysis Across Independent Studies

This thesis develops statistical tools to assess directional replicability—whether at least r of n independent studies exhibit effects with the same sign. Chapter 1 establishes foundational hypothesis testing principles, including error rates and multiplicity control. Chapter 2 formalizes replicability using the partial conjunction (PC) framework and introduces AdaFilter, an adaptive method that filters hypotheses to improve power while maintaining error control. Chapter 3 extends AdaFilter to directional settings, implementing a two-pipeline approach (positive/negative evidence) with level splitting to preserve familywise error rate (FWER) control. Simulations demonstrate the method’s ability to balance sensitivity and replicability thresholds, with results interpreted under independence assumptions. The work underscores the need for rigorous, sign-consistent replicability claims in high-throughput research