| dc.description.abstract | Interest in using biomarkers for prognosis, diagnosis, and screening continues to grow in many clinical areas. However, most biomarkers have only modest predictive capacity and therefore are not clinically useful. As the cost of measuring individual biomarkers declines, investigators are increasingly interested in combining biomarkers to create tools that are clinically useful. Creating such tools involves constructing a combination of a set of biomarkers and evaluating its predictive capacity; in some settings where a large number of biomarkers are available, combination selection may be necessary. In this dissertation, we consider particular challenges that may arise in the construction, evaluation, and selection of biomarker combinations and propose methods to address these challenges. We first propose a distribution-free method to construct biomarker combinations by maximizing the true positive rate while constraining the false positive rate at some clinically acceptable level. We also consider the potential role of multilevel outcomes in combination construction and selection when there is interest in predicting a particular level of the outcome due to its clinical importance. Finally, we address issues related to the use of biomarker data from multiple centers. We describe the potential role of center in these studies, demonstrate problems with currently used methods for constructing biomarker combinations, present appropriate likelihood-based methods for constructing combinations, and consider how to correctly evaluate the performance of combinations in this setting. We then move beyond the maximum likelihood framework and propose a method that directly maximizes a center-adjusted measure of performance while allowing for penalization of variability in performance across centers. This research provides investigators with novel insights and methods that will facilitate the development of biomarker combinations for diagnosis, prognosis, and screening. | |
