Structure–Adaptive Sequential Testing for Online False Discovery Rate Control

Consider the online testing of a stream hypotheses where real-time decision must be made before next data point arrives. The error rate is required to controlled at all points. Conventional simultaneous rules are no longer applicable due more stringent constraints and absence future data. Moreover, decision-making process may come halt when total budget, or alpha-wealth, exhausted. This work develops new class structure-adaptive sequential (SAST) for false discovery (FDR) control. A key element in our proposal alpha-investing algorithm that precisely characterizes gains losses making. SAST captures time varying structures stream, learns optimal threshold adaptively an ongoing manner optimizes alpha-wealth allocation across different periods. We present theory numerical results show asymptotically valid FDR control achieves substantial power gain over existing rules.