Title:Controlling the FDR in variable selection via multiple knockoffs

Abstract:Barber and Candes recently introduced a feature selection method called knockoff+ that controls the false discovery rate (FDR) among the selected features in the classical linear regression problem. Knockoff+ uses the competition between the original features and artificially created knockoff features to control the FDR [1]. We generalize Barber and Candes' knockoff construction to generate multiple knockoffs and use those in conjunction with a recently developed general framework for multiple competition-based FDR control [9].
We prove that using our initial multiple-knockoff construction the combined procedure rigorously controls the FDR in the finite sample setting. Because this construction has a somewhat limited utility we introduce a heuristic we call "batching" which significantly improves the power of our multiple-knockoff procedures.
Finally, we combine the batched knockoffs with a new context-dependent resampling scheme that replaces the generic resampling scheme used in the general multiple-competition setup. We show using simulations that the resulting "multi-knockoff-select" procedure empirically controls the FDR in the finite setting of the variable selection problem while often delivering substantially more power than knockoff+.