|
Abstract:
|
We propose a novel framework of sufficient variable screening for ultrahigh dimensional semi-competing risks data based on a class of powerful independence measures proposed recently. Compared with existing sure independence screening methods for survival data, which only consider marginal relationship between the survival estimates and each predictor, our approach further incorporates joint information among predictors to refine the screening. Furthermore, our procedure takes advantage of the independence measures and avoid estimating the survival function. We establish the sure screening property of the proposed approach in the regime of sufficient variable selection. A partial sufficient variable screening procedure conditioning on a known set of important predictors is also introduced. Numerical studies across a variety of regression and classification settings demonstrate the superiority of our method.
|
