Model Selection Strategy for Cox Proportional Hazards Model

Abstract

Often in survival regression modelling, not all predictors are relevant to the outcome variable. Discarding such irrelevant variables is very crucial in model selection. In this research, under Cox Proportional Hazards (PH) model we study different model selection criteria including Stepwise selection, Least Absolute Shrinkage and Selection Operator (LASSO), Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and the extended versions of AIC and BIC to the Cox model. The simulation study shows that varying censoring proportions and correlation coefficients among the covariates have great impact on the performances of the criteria to identify a true model. In the presence of high correlation among the covariates, the success rate for identifying the true model is higher for LASSO compared to other criteria. The extended version of BIC always shows better result than the traditional BIC. We have also applied these techniques to real world data.