Evaluating prediction performance of longitudinal biomarkers under cohort and two-phase study designs

Abstract

Risk prediction and evaluation of predictions based on longitudinal biomarkers are of interest in treatment selection, preventive medicine and management of chronic diseases. Methods to evaluate risk predictions in a longitudinal setting are limited to the area under the receiver operating characteristic curves and prediction error. In this dissertation, we evaluate two approaches to risk prediction in the longitudinal setting: joint modeling and partly conditional modeling. We develop estimation procedures for more flexible and robust partly conditional models, demonstrate their adaptability and applicability, and provide a smoothing technique to account for measurement error in marker data. We develop nonparametric estimators of clinically relevant measures of prediction quality in the longitudinal setting under cohort, case-cohort, stratified case-cohort and nested case-control study designs. We provide resampling-based inference procedures for all estimators under the four study designs. We evaluate our methods using simulation studies and illustrate them on the End Stage Renal Disease Study dataset and a nested case-control study within the HALT-C clinical trial.