Methods for Risk Markers that Incorporate Clinical Utility

Abstract

Risk markers are often used to help make clinical decisions. In this dissertation, we focus on developing statistical methods that account for the utility of a risk marker. We address problems of individualized decision-making, calibration, and combining multiple biomarkers when the ultimate goal is to use the combination for clinical decision-making. We review methods of estimating clinical utility from Bayesian and frequentist standpoints and draw connections between the two frameworks. We additionally consider the appropriateness of each framework to the individual decision-making problem. When existing risk models are applied to new populations, issues of miscalibration can arise. We propose two methods for recalibration that account for the clinical context in which the risk model will be used. Finally, we address the problem of combining risk markers into a single “composite” biomarker. We present a non-parametric method for developing linear combinations of risk markers that maximizes net benefit. We evaluate our methods using simulation studies and apply them to data from prostate cancer, cardiac disease, and diabetes studies.