Variable selection by ensembles for the Cox model

Abstract

Using Cox regression as the main platform, we study the ensemble approach for variable selection. We use a popular real-data example as well as simulated data with various censoring levels to illustrate the usefulness of the ensemble approach, and study the nature of these ensembles in terms of their strength and diversity. By relating these characteristics to the ensemble's selection accuracy, we provide useful insights for how to choose among different ensemble strategies, as well as guidelines for thinking about how to design more effective ensembles.

Keywords:

• bootstrap
• Cox regression
• ensemble learning
• survival analysis
• strength-diversity trade-off
• variable selection

Acknowledgements

This work is partially supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada.