In medical research, receiver operating characteristic (ROC) analysis is a widely used tool for evaluating the performance of continuous diagnostic tests and relies on a gold standard test to determine subjects’ true disease status. However, in clinical practice, not all subjects who undergo a diagnostic test have their disease status verified because of ethical, financial, or clinical constraints. Since this missingness is often informative rather than completely random, restricting analysis to only those with verified disease status can introduce verification bias that undermines the evaluation of the diagnostic test’s performance.

This dissertation develops statistical methods for evaluating continuous diagnostic tests when true disease status is missing at random. Specifically, we propose and investigate several interval estimation methods for three key ROC-based measures of diagnostic accuracy: the Youden index, the area under the ROC curve, and sensitivity at a fixed level of specificity. Verification bias is addressed through imputation and reweighting techniques, and confidence intervals are constructed using bootstrap resampling and empirical likelihood methods. Extensive simulation studies and real data applications demonstrate the effectiveness and robustness of the proposed approaches across a range of clinically relevant settings. To facilitate practical implementation, an R package providing the proposed methods is also developed.