A Variable Selection Method for High-Dimensional Survival Data

Survival data with high-dimensional predictors are regularly collected in many studies. Models with a very large number of covariates are both infeasible to fit and likely to incur low predictability due to overfitting. The selection of significant variables plays a crucial role in estimating models. Even if several approaches that identify variables in presence of censored data are available in literature, there is not unanimous consensus on which method outperforms the others. Nonetheless, it is possible to exploit the advantages of methods to get the final set of covariates as good as possible. Therefore, we propose a method that combines different variable selection procedures by using the subsampling technique, for identifying as relevant those covariates that are selected most frequently by the different variable selectors on subsampled data. By a simulation study, we evaluate the performance of the proposed procedure and compare it with other techniques.