Title
Estimating Diagnostic Accuracy From Designs With No Gold Standard Evaluation and With Partial Gold Standard Evaluation
Speaker
Paul Albert, Biometric Research Branch, National Cancer Institute
Abstract
Interest often focuses on estimating sensitivity and specificity of a group of raters or a set of new diagnostic tests in situations where gold standard evaluation is invasive or expensive. In a typical situation a group of raters or a series of diagnostic tests assess disease status on a group of individuals. For situations in which no gold standard evaluation is available, various authors have proposed latent class modeling approaches for estimating diagnostic error and prevalence. We discuss a potential problem with these approaches. Namely, we show that when the conditional dependence between tests is misspecified, estimates of sensitivity, specificity and prevalence can be severely biased. Importantly, we demonstrate that with a small numbers of tests, likelihood comparisons and other model diagnostics may not be able to distinguish between models with different dependence structures. While these results caution against using these latent class models, the difficulties of obtaining gold standard verification remain a practical reality. We provide a compromise in which gold standard information is collected on a subset of subjects. We propose both semi-latent class models and a multiple imputation approach for estimating diagnostic error and prevalence with partial gold standard evaluation. Through analytic work and simulations, we show that even with a small percentage of verified individuals, in most cases, these approaches are substantially more robust than latent class models without any gold standard information. We illustrate our methodological work with data analyses from two medical examples. This work is joint with Lori Dodd at NCI.
Meet the speaker in Room 212 Cockins Hall at 4:30 p.m. Refreshments will be served.
