Improving Stepwise Proportional Hazards Regression Using a SAS Macro

Abstract

Stepwise covariate selection is a popular method for multivariable regression model building. A model with fewer covariates is numerically stable and easier to use. Based on the different significance levels pre-specified by statisticians, different covariates are included in the model. Further analyses with these models might introduce biases. This paper proposes a novel method to select covariates for stepwise proportional hazards regression without pre-setting a significance level. Multiple models containing different numbers of covariates were outputted for final model selection. A user-oriented SAS macro was developed. Users of the macro may determine the final models, based on estimated characteristic changes of the overall models, the variations of the covariate effects on the response variable and their special needs. With this method, model selections are much easier than with purposeful or the best subsets method. This method improved stepwise covariate selection processes. Broad applications are expected.