Nonparametric estimation of ROC curves based on Bayesian models when the true disease state is unknown

We develop a Bayesian methodology for nonparametric estimation of ROC curves used for evaluation of the accuracy of a diagnostic procedure. We consider the situation where there is no perfect reference test, that is, no “gold standard”. The method is based on a multinomial model for the joint distribution of test-positive and test-negative observations. We use a Bayesian approach which assures the natural monotonicity property of the resulting ROC curve estimate. MCMC methods are used to compute the posterior estimates of the sensitivities and specificities that provide the basis for inference concerning the accuracy of the diagnostic procedure. Because there is no gold standard, identifiability requires that the data come from at least two populations with different prevalences. No assumption is needed concerning the shape of the distributions of test values of the diseased and non diseased in these populations. We discuss an application to an analysis of ELISA scores in the diagnostic testing of paratuberculosis (Johne’s Disease) for several herds of dairy cows and compare the results to those obtained from some previously proposed methods.