American Statistical Association New York City Metropolitan Area Chapter Mailman School of Public Health Columbia University Department of Biostatistics Colloquium

Accurate diagnosis of disease in many fields such as Alzheimer's disease (AD) is the first step towards its control. For example, cerebrospinal fluid (CSF) tau is useful in detecting AD. CSF tau is often assumed to measure the underlying true biological quantity without any error when its sensitivity and specificity are calculated. However, in practice this assumption almost never holds. Typically, the observed sensitivity and specificity are biased when the diagnostic test is subject to measurement error. When an internal reliability sample is available for CSF tau, we have shown that averaging replicates and ignoring measurement error can sometime lead to serious biases of sensitivity and specificity estimates. We then present a bias-correction approach to remove the biases of the sensitivity and specificity estimates introduced by the measurement error based on an internal reliability sample. Asymptotic distributions were obtained for the proposed estimators. Extensive simulations were conducted to evaluate the performance of the proposed bias-correction approach. All methods are illustrated using the ongoing University of Pennsylvania AD biomarker study.