Online False Discovery Rate Control with Dependent Data Streams

Abstract

Online false discovery rate (FDR) control addresses multiple testing problems where hypotheses arrive sequentially and decisions must be made without access to future data. In this talk, we begin by introducing the basic framework of online FDR control and discussing challenges that arise when hypotheses are generated from dependent data streams.
We then consider a statistical Sequential Inference System (SIS) model that describes latent state dynamics underlying the hypothesis sequence. Within this framework, we discuss how the existing offline multiple testing procedures can be leveraged to construct online FDR control methods. In particular, we present a general perspective on converting offline FDR procedures to sequential settings, and illustrate how the local index of significance (LIS) enjoys special structural properties that make it particularly suitable for such offline-to-online constructions.

線上視訊請點選連結