Evaluation of the Accuracy of Diagnostic Tests From Repeated Measurements Without a Gold Standard

摘要

A model is presented to evaluate the accuracy of diagnostic tests from data from individuals that are repeatedly tested in time. Repeated measurements from three diagnostic tests for foot-and-mouth disease, applied to vaccinated and experimentally infected cattle, were analyzed. At any time the true disease status of the individuals was unknown, i.e., no gold standard was available. The model allows for correlation between repeated test results, in consequence of the underlying structure for the unknown true disease status, but also by the distribution of the test results conditional upon true disease status. The model also allows for dependence between the different diagnostic tests conditional upon true disease status. Prior information about the structure of the prevalence and the specificity of the tests was incorporated in a Bayesian analysis. Posterior inference was carried out with Markov chain Monte Carlo. Simulated data were analyzed to gain insight into the performance of the posterior Bayesian inference. The simulated data are typical for the expensive and, therefore, modestly sized infection experiments that are conducted under controlled conditions.