Abstract Summary/Description
In medical diagnostic, especially when the gold standard test is costly or invasive, it is a common case that not all subjects with diagnostic test results ultimately have their true disease status verified. This generally leads to a verification biased evaluation of the ability of the diagnostic tests. To address this problem, Todd A. Alonzo and Margaret S. Pepe introduced four partially parametric point estimators of sensitivity, specificity and ROC under the missing-at-random (MAR) assumption. However, to the best of our knowledge, no verification bias correction interval estimation for continuous tests has been specifically developed yet. This paper aims to fill this gap. Inspired by bootstrap method and the method of variance estimates recovery (MOVER), we propose new bias-corrected interval estimations for the two-class Youden index, which is a well-accepted measure of accuracy for continuous tests, under the MAR assumption. Extensive Monte Carlo simulation studies are conducted to test and compare our approach. The results show that when 36% of the subjects have their true disease status missing, our proposed methods can achieve an expected 95% coverage probability with appropriate length. Real data study using Wisconsin Diagnostic Breast Cancer biomarker data also confirms our approaches' strong robustness across all scenarios, even when only 36% of patients have verified diagnoses and without any assumptions about the underlying disease model.