Feature screening for high dimensional data with response and covariates missing at random

Feature screening has been proved to be effective and fashionable method for analyzing high-dimensional data. It is a challenge in identifying important feature for the high-dimensional data when both response and covariates are missing at random. In this paper, we consider a feature screening procedure based on distance correlation between the marginal distribution of the covariate and the marginal distribution of the response for missing data, where the marginal distribution of missing variable is estimated by inverse probability weighting method. We also propose a computationally very efficient algorithm based on proposed screening procedure. The results of some numerical experiments show that the proposed method and algorithm have a better performance for linear model even if both response and covariates missing at random, and also work well for nonlinear model.