Abstract: We develop two sufficient variable screening procedures utilizing the newly proposed censored distance correlation measures for ultrahigh-dimensional right censored data. Compared to many existing methods, our procedures more effectively detect active predictors that are marginally independent of the response. They are also model-free and robust against model misspecification. Through simulations and real data analysis, we demonstrate the distinct advantages of our proposed procedures over existing variable screening methods.

Keywords words and phrases: Distance correlation, feature screening, independence measure.