B0200
Title: Doubly robust evaluation of receiver operating characteristic under covariate shift with high dimensional features
Authors: Molei Liu - Harvard T.H. Chan School of Public Health (United States) [presenting]
Tianxi Cai - Harvard School of Public Health (United States)
Abstract: Transfer learning plays an important role in the presence of covariate shift. Most works in this regime focus on model estimation, while robust and efficient model accuracy evaluation on the target population lacks enough attention despite its importance. We tackle this problem through a novel augmented estimation approach for the receiver operating characteristic parameters. The proposed estimators are doubly robust in the sense that it is root-n consistent when one correctly specifies at least one of the two nuisance models: a density ratio model characterizing the covariate shift and an imputation model for the response $Y$. Our method targets a low dimensional outcome model. It accommodates high dimensional shifted features by calibrating the estimating equations for the nuisance models to correct for their estimation (regularization) bias under high dimensionality.
