Ensemble Learning Methods for Survival Prediction in Studies with Outcome-Dependent Sampling Designs

Outcome-dependent sampling designs are widely used in biomedical research as cost-effective alternatives to full cohort studies. However, much of the existing research in survival analysis focuses on estimation and inference for semiparametric models, with limited development in survival prediction methods. This dissertation addresses this gap by developing ensemble learning approaches tailored for survival prediction in studies utilizing outcome-dependent sampling designs. In the first project, we propose a super learner algorithm for survival prediction in case-cohort and generalized case-cohort studies. The algorithm achieves asymptotic model selection consistency and uniform consistency while exhibiting superior finite-sample performance. Simulation studies demonstrate improved prediction accuracy compared to super learners trained on simple random samples of equal size. The second project extends the random survival forest (RSF) framework to case-cohort and generalized case-cohort designs by incorporating multiple imputation by chained equations (RSF-MICE) and substantive model compatible imputation (RSF-SMCI). These approaches enhance prediction accuracy by leveraging inexpensive covariate data from the full cohort, facilitating better use of available information. In the third project, we address scenarios where full-cohort covariate data are not predictive of survival outcomes. We develop a modified RSF method that employs a resampling scheme preserving the case-cohort data structure, along with tailored tree-splitting and within-node estimation techniques. The proposed method achieves uniform consistency and demonstrates strong performance in finite samples. The final project introduces a gradient boosting machine designed for general failure-time outcome-dependent sampling. This approach incorporates a weighting scheme to correct for sampling bias and improve predictive accuracy. A weighted empirical risk based on the Brier score is proposed for hyperparameter tuning. The method shows robust performance across various simulation settings.