报告时间：2016年12月16日（周五）上午8:30

报告地点： 京师学堂第七会议室

报 告 人： 崔欣萍，美国加州大学河滨分校统计系正教授、系主任

报告题目：Sequential Multiple Testing for Variable Selection in High Dimension-al Linear Models

报告摘要：Covariance test (Lockhart et al. 2014) provided p-values for all variables that enter into a linear model sequentially along a lasso solution path. Using these p-values to select a model with inferential guarantees is equivalent to multiple hypothesis testing setting where the hypotheses are ordered. In this paper, we proposed a sequential multiple hypothesis testing framework, which considers multiple testing within each step and across all steps along the lasso solution path. Under this framework, we designed stepwise p-values and applied Benjamini-Liu (BL, 1999) step down procedure. We compare it with the Single Step-Down method and ForwardStop and StrongStop procedures. Simulation studies show that our proposed procedure has higher power with FDR controlled at the desired level, especially for large scale and high-dimensional data. Diabetes study and framingham heart study examples are also presented.