Discovery of Gene-Specific Time Effects on Survival

The success of variable screening of time-to-event data can be enhanced by identifying variables which influence the timing of mortality. However, most existing screening approaches based on a uniform screening criterion may fail to capture such subtle survival differences especially at the early or late event time due to the imposed assumptions on the survival function. By artificially inflating the difference in survival functions across strata defined by each candidate covariate, we propose a general class of screening criteria that can detect biomarkers with early or delayed impact on survival distributions, providing new perspectives for high-dimensional survival analysis.