Title: A double application of the Benjamini-Hochberg procedure and its refinement
Authors: Qingyun Cai - Xiamen University (China) [presenting]
Abstract: The Benjamini-Hochberg (BH) procedure controls the false discovery rate (FDR), and optimizes signal discovery especially in large dataset. However it does not consider any index information of the null hypotheses. A double application of the BH procedure on two-level hierarchical datasets is proposed. The first application is to identify p-value batches; the second application is to identify null hypotheses rejections in each batch. It is shown that the double application not only maintains the power of BH, but also satisfies an average FDR control and reduces FDR when the signals are clustered. Based on this, a refined procedure of the double application is proposed by incorporating proportion of false null hypotheses to further improve performance.