Assessing accuracy of a continuous medical diagnostic or screening test in the presence of verification bias

Abstract

In studies to assess accuracy of a medical diagnostic or screening test, often definitive disease assessment is too invasive or expensive to be ascertained in all subjects. It makes practical and ethical sense to select a higher percentage of subjects with a positive screening test for disease assessment than patients with a negative screening test. However, analyses using only the results for this subset of subjects can cause biased accuracy estimates. This is known as verification bias or work-up bias.Many screening tests, such as tests based on serum concentrations or biomarkers, yield continuous results; however, available bias-correction methods can only accommodate tests with discrete responses. We propose new verification bias-correction methods to assess accuracy of a continuous test based on (1) an extension of the work of Begg and Greenes (Biometrics, 1983) for discrete tests and (2) mean score, (3) inverse probability weighting, and (4) semi-parametric efficient methods. Asymptotic distribution theory is developed for estimators of disease prevalence, true positive rate, and false positive rate. The different methods are compared using theoretical and simulation results and applied to data from a newborn hearing screening study.