Title: A variable selection method for high-dimensional survival data
Authors: Sara Milito - University of Salerno (Italy) [presenting]
Marialuisa Restaino - University of Salerno (Italy)
Francesco Giordano - University of Salerno (Italy)
Abstract: In many studies, survival data with high-dimensional predictors are regularly collected. 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 and it is particularly difficult in high-dimensional settings, where the number of covariates may be greater than the sample size ($n <p$ and $n << p$). Several approaches select variables in presence of censored data are available in literature, but there is not unanimous consensus on which method outperforms the others. However, it is possible to exploit the advantages of all methods to obtain the final set of covariates as good as possible. Therefore, in order to improve the performance of variable selection methods, we propose a method that combines different procedures with subsampling, for identifying as relevant those covariates that are selected most frequently by the different variable selection methods on the subsampled data. By a simulation study, we evaluate the performance of the proposed procedure and compare it with other techniques.