Selecting optimal thresholds in ROC analysis with clustered data

We tackle estimation of receiver operating characteristic (ROC) surfaces and selection of the optimal pair of thresholds for continuous diagnostic tests given covariates in clustered data when the disease status is described by three ordinal classes. The approach is based on a linear mixed-effect model which accounts for both the clusters and the covariates effects. The asymptotic properties of estimators are studied. Simulation studies are performed to assess the performance of estimators.