Author: Feng Miao
Date: 6/24/2022
Executive Summary:
To translate a continuous or ordinal diagnostic variable into a clinical decision, it is necessary to determine a cutoff point that dichotomizes the patients into distinct groups for risk stratification, disease diagnosis, and management of patient care. There are numerous existing criteria only for establishing optimal cut-points of unmatched or unpaired data. Furthermore, those methods rely on the assumption that the diagnostic variable is heterogeneously distributed between two groups.In the paired case-control design, data matching reduces the bias due to confounding factors or covariates. The risk of type 1 diabetes (T1D) involves both genetic and non-genetic factors. TrialNET clinical data include biomarkers from autoantibody-positive and autoantibody-negative, family-matched siblings. Conditional ROC curves have been developed to appropriately accommodate correlated biomarkers arising from this matched case-control design. However, the approach for optimizing cut-points in matched settings has not been investigated. In this work, correlated biomarker data are transformed into an uncorrelated structure, then optimal cut-points are determined by traditional non-parametrical and parametrical approaches.The optimized cut-point separates the two groups with the least misclassification. When the independence assumption is violated， thebiomarker dichotomized by the optimal cut-point efficiently rejects the null hypothesis. A small p-value from the hypothesis test indicates a potential predictive biomarker. In this research, the new hypergeometric distributions and exact tests overcome the conservative nature of Fisher’s exact test.
