A0568
Title: Variable selection in Cox model with partially observed covariates under missing at random
Authors: Ujjwal Das - Indian Institute of Management Udaipur (India) [presenting]
Abstract: The focus is on the performances of several feature selection approaches in the Cox proportional hazards model with the covariates, which are partially observed according to a missing at-random mechanism. Two multiple imputation methods are adopted: (i) Predictive mean matching using the chained equations model and (ii) bootstrap-based nonparametric methods to impute missing observations. The standard approach to combining analyses of imputed data sets commonly yields different sets of variables. The imputed data sets are stacked, and the integrated data is analyzed. Five variable selection methods are considered. The first three methods are based on horizontal stacking of imputed datasets and identify variables using: (a) Group LASSO, (b) group SCAD, and (c) group MCP, employing the corrected Akaike information criterion (AIC) to select the penalizing parameter. The other two methods are based on vertical stacking of imputed datasets and select the predominant variables using some likelihood ratio tests proposed in a recent study. The proposed methods are investigated numerically. Finally, the methods are illustrated on data from a real-world oncology experiment.
