The Fall 2023 SDS Seminar Series continues on September 22nd from 2:00 p.m. to 3:00 p.m. with Will Fithian (Department of Statistics, University of California, Berkley). This event is in-person.
Title: Estimating the False Discovery Rate of Model Selection
Abstract: I will introduce a novel method for assessing the false discovery rate (FDR) of a model selection method such as the LASSO, forward-stepwise regression, or the graphical lasso. Our method gives an FDR estimate whose bias is provably non-negative, and can be employed alongside methods like cross-validation or SURE to trade off model selection accuracy against prediction error, or simply to warn the analyst about the likelihood that an estimator's FDR is large. I will also present the main technical tool for this method, conditional calibration, which can also be employed to adjust the Benjamini-Hochberg procedure for dependence, or to improve the power of knockoff methods. This is joint work with Lihua Lei and Yixiang Luo.