Online control of the false discovery rate in biomedical research

Modern biomedical research frequently involves testing multiple related hypotheses, while maintaining control over a suitable error rate. In many applications the false discovery rate (FDR), which is the expected proportion of false positives among the rejected hypotheses, has become the standard error criterion. Procedures that control the FDR, such as the well-known Benjamini-Hochberg procedure, assume that all p-values are available to be tested at a single time point. However, this ignores the sequential nature of many biomedical experiments, where a sequence of hypotheses is tested without having access to future p-values or even the number of hypotheses. Recently, the first procedures that control the FDR in this online manner have been proposed by Javanmard and Montanari (Ann. Stat. 2018), and built upon by Ramdas et al. (NIPS 2017, ICML 2018). In this paper, we compare and contrast these proposed procedures, with a particular focus on the setting where the p-values are dependent. We also propose a simple modification of the procedures for when there is an upper bound on the number of hypotheses to be tested. Using comprehensive simulation scenarios and case studies, we provide recommendations for which procedures to use in practice for online FDR control.
Comment: 48 pages, 18 figures; Updated references