Feature selection in Cox model with partially observed covariates: Application to oncology trials

In many real-life experiments with human subjects, missing data are common. Multiple imputation is widely used to handle unobserved data points. In statistical research, selecting important variables from multiple imputed datasets can be challenging, as each imputed data set may yield different sets of variables. Over the last decade, stacking imputed datasets and analyzing the resulting integrated data has gained attention. In this article, we consider both horizontal and vertical stacking approaches. The horizontal stacking approach in conjunction with different group penalties is discussed alongside the recently proposed vertical appending method, for identifying predominant variables under time-to-event data. The proposed methods are investigated numerically. Finally, the methods are illustrated in two real-world oncology experiments.