Bayesian modeling of ROC curves for bovine paratuberculosis ELISA tests in the absence of a gold standard

摘要

Receiver operating characteristic (ROC) curves plot the sensitivity versus (1-specificity) over all cut-off points of a quantitative diagnostic test and therefore they provide a cut-off independent measure of test accuracy. Estimation of ROC curves (and the associated area under the curve) is straightforward when a gold-standard reference test is available but methods to estimate curves in the absence of a gold standard have been inadequately studied. We developed a Bayesian ROC analysis that can be applied to multiple correlated diagnostic tests with and without a gold standard. Simulation studies showed that the discrimination ability of the no-gold standard (NGS) method was adequate compared with the gold standard (GS) method providing that the overlap between the 2 distributions of ELISA values was not too great. We used the proposed method to analyze results of 2 serum ELISAs and fecal culture for bovine paratuberculosis. Data were available from 449 cattle with a positive fecal culture in paratuberculosis-infected herds and from 393 cattle from paratuberculosis-free herds. Fecal culture results were the averaged values for 3 laboratories, where scores of 1, 2, 3, and 4 represented HEY colony counts of 1-9, 10-49, 50-99 and >= 100 colonies per slant, respectively. Log transformation of ELISA S/P ratios was necessary to achieve bivariate normality, which is a necessary assumption for a parametric analysis. The Parachek ELISA had a greater area under the ROC curve than the HerdChek ELISA by both GS and NGS methods. The NGS method provided adequate discrimination only when the subset of infected cattle with fecal culture scores of >= 3 was used. We are investigating semi-parametric extensions of the model to allow for situations where a suitable data transformation cannot be found.