On spike and slab empirical Bayes multiple testing (Q2215749)

The authors consider multiple testing of hypotheses, a topic related to identifying the active variables among a large number of candidates in high-dimensional statistical models. High-dimensional data typically involve more than thousands of variables with only a small part of them being significant. Bayesian multiple testing methodology is applied where the testing is based on comparing posterior probabilities of the hypotheses under consideration. The choice of prior relies on empirical Bayes approaches that aim at calibrating the prior in a fully automatic, data-driven, way. The paper explores a connection between empirical Bayes posterior distributions and false discovery rate (FDR) control. In the Gaussian sequence model, the results show that spike and slab priors produce posterior distributions with particularly suitable multiple testing properties. The authors demonstrate that a uniform control is possible up to a constant term away from the target control level. This constant is very close to \(1\) in simulations, and can even be shown to be \(1\) asymptotically for some subclass of sparse vectors. In this way, a theoretical validation is provided for the common practical use of posterior-based quantities in frequentist FDR control. The theoretical results are illustrated with numerical experiments.