The differential diagnosis of a disease is often based on the information obtained from multiple diagnostic tests or multiple applications of the same test. The usual assumption in such situations is that the test results are statistically independent within each subject conditional on knowing the true disease status. This assumption greatly simplifies the statistical analysis of such data. In practice, however this assumption may be violated, as for example when there is a certain subject-related characteristic that may increase or decrease the probability of detection in two or more tests. The classical or frequentist solutions that account for the correlation between tests require a minimum of four different tests to obtain an identifiable solution. However, it is not always possible to have results from four different tests, particularly when tests are expensive, time consuming or invasive. Our objective in this thesis is to draw simultaneous inferences about the prevalence and test parameters while adjusting for the possibility of conditional dependence between tests, particularly in the situation when we have three or fewer tests, leading to a non-identifiable problem. We do so by way of a Bayesian approach, which utilizes available information about the prevalence and test parameters summarized in the form of prior distributions. The first of the two methods we propose models the dependence as a direct effect between each pair of tests. The second method uses a random effects model and simulates the dependence between tests via their sensitivity and specificity which are modeled as functions of a latent, subject-specific 'disease intensity'. Both models are based on dichotomous tests and the parameters are estimated using a Gibbs Sampler. It was found that ignoring the conditional dependence between tests could lead to misleading estimates of the sensitivities and specificities of the tests and of disease prevalence. Therefore, the methods presented here m
展开▼
