Title: Weeding out early false discoveries along the Lasso Path with knockoffs
Authors: Malgorzata Bogdan - University of Wroclaw (Poland) [presenting]
Weijie Su - University of Pennsylvania (United States)
Emmanuel Candes - Stanford (United States)
Asaf Weinstein - Stanford University (United States)
Abstract: It is now widely recognized that the popular Lasso method of identfying predictors in large data bases often suffers from including large number of false discoveries. In a recent work this phenomenon has been quantitatively described using the framework of the Approximate Message Passing (AMP) theory. Specifically, it was shown that the Lasso is limited by the FDR-power tradeoff, which in case of moderately dense signals does not allow to simultaneously obtain high power and small false discovery rate. We will use AMP theory to show that this limitation can be overcome by combining Lasso with the recent method of knock-offs, for controlling FDR in the context of multiple regression.