Presentation: Variable Screening using Non-Linear Association Measures
Speaker: Adam Elder, Graduate Student, UW Biostatistics
Abstract: Suppose one observes a collection of covariates and an outcome, and wishes to test whether the outcome is associated with any of the covariates, as quantified by a user-specified association measure. We develop a method to test this sharp null using, as test statistic, the L_p norm of the vector of covariate-specific association estimates. For any fixed p, we propose a test with valid type I error control whenever asymptotically linear estimators of the association measures are used. We are currently developing a method to data-adaptively determine the value of p that maximizes power while maintaining type I error control.