Title: Model-free multiple testing using mirror statistics (MMM)
Authors: Zhigen Zhao - Temple University (United States) [presenting]
Xin Xing - Virginia Tech University (United States)
Abstract: The general regression analysis is considered, and the relation between a univariate response and a p-dimensional covariate is studied. We assume the general multi-index model with an unknown link function. It is assumed that the response depends on the covariate via some linear combinations, which is characterized by the central subspace. For all the covariates, we want to test the hypothesis of whether each individual predictor plays any role in the central subspace subject to the control of the false discovery rate. We combine the method of sufficient dimension reduction and the Gaussian mirror to construct the MMM method, standing for Model-free Multiple Testing using Mirror Statistics. It is shown that MMM controls the FDR at the desired level asymptotically. Numerical evidence has shown that MMM is much more powerful than all its alternatives.