B0612
Title: FDR estimation for variable selection methods
Authors: Yixiang Luo - UC Berkeley (United States) [presenting]
William Fithian - University of California at Berkeley (United States)
Lihua Lei - Stanford University (United States)
Abstract: In variable selection, whether the selected variables are truly relevant to the outcome is a natural concern in many applications. A framework is proposed to assess the false discovery rate (FDR) for a large family of variable selection procedures, including Lasso and forward stepwise selection in the Gaussian linear model and graphical Lasso in the Gaussian graphical model. The FDR estimator has a non-negative bias. And it has vanishing variance under certain conditions in the Gaussian linear model. Practical examples with real data are given for Lasso and graphical Lasso.
