Abstract
Many areas of research involve the analysis of correlated biomarker data. Correlated biomarker data arise from study designs including subjects that contain shared genetic or environmental factors. In a familial-matched design, the information about correlation might help to identify family members at increased risk of disease development and may lead to initiating treatment to slow or stop disease progression. Receiver operating characteristic (ROC) curves have been widely used in medical research to illustrate the performance of a biomarker in correctly distinguishing between diseased and non-diseased groups. Approaches appropriate to a familial-matched case-control design should accommodate inherent correlation in correctly estimating the biomarker’s ability to differentiate between groups, as well as handle estimation from a matched case-control design. It is sensible to expect biomarkers to demonstrate improved ability to partition between groups from a paired case-control design. There are not currently any known methods where the estimated performance based on ROC construction can reflect improved classification performance under a paired design. This dissertation will first review some developed methods for ROC curve estimation and will discuss the limitations and gaps of current methods for analyzing correlated familial paired data. A new approach using conditional ROC curves will be demonstrated to provide ROC curves for correlated paired data. The proposed approach will use the information about correlation among biomarker values, producing conditional ROC curves that evaluate the ability of a biomarker to discriminate between diseased and non-diseased subjects in a familial-paired design. Also, some extensions to this approach as well as an alternative approach to conditional ROC curve estimation for family paired case-control data are presented.