B1341
Title: Feature selection for competing risks model in high and ultra-high dimensions
Authors: Marialuisa Restaino - University of Salerno (Italy) [presenting]
Francesco Giordano - University of Salerno (Italy)
Abstract: In the analysis of time-to-event data, competing risks data are encountered when individuals may fail from multiple causes and the occurrence of one failure event precludes the others from happening. Two main approaches can be used, to investigate the effects of covariates on the hazard function: cause-specific hazard (CSH) model and subdistribution hazard (SDH) model. The difference between these two approaches relies on the definition of the risk set. In CSH, subjects who experience the competing events are treated as censored, while in SDH they are included in the risk set. In many applications involving competing risks, identifying variables that have effects on CSH or SDH is a critical task. In the CSH model, screening and variable selection methods developed for Cox model can be easily extended. For the SDH approach, due to the different definitions of the risk set, naive applications of these procedures may be problematic and not suitable. This is particularly true when the number of covariates is larger than the number of observations ($n<p$ and $n<<p$) and in presence of multicollinearity between covariates. Thus, the aim is to compare the performance of some existing methods for screening and selecting the most significant variables, for both CSH and SDH models, for highlighting their main advantages and disadvantages and proposing a new procedure able to identify the relevant covariates in the framework of high and ultra-high dimensions.