A Modified Random Survival Forest for Improving Prediction Accuracy in Case‐Cohort and Generalized Case‐Cohort Studies

Case‐cohort and generalized case‐cohort study designs offer cost‐effective alternatives to full cohort studies for analyzing associations between risk factors and survival outcomes. However, existing methods for nonparametric prediction using random survival forest either ignore the sampling design, leading to biased estimates, or incorporate inverse probability weights in the bootstrapping without fully leveraging the estimation efficiency of the case‐cohort or generalized case‐cohort study design. In this paper, we propose a modified random survival forest to improve prediction accuracy under the above designs. We enhance the prediction accuracy by proposing a resampling scheme that helps to retain the case‐cohort structure and using the tree splitting and within‐node estimation methods tailored for the study design. We show that our approach is uniformly consistent, and the simulation study demonstrates satisfactory finite sample performances. We apply our method to predict the incidence of diabetes using data from a generalized case‐cohort study conducted within the Atherosclerosis Risk in Communities Study.