Bayesian analysis of diagnostic test accuracy when disease state is unverified for some subjects.

Studies of the accuracy of medical tests to diagnose the presence or absence of disease can suffer from an inability to verify the true disease state in everyone. When verification is missing at random (MAR), the missing data mechanism can be ignored in likelihood-based inference. However, this assumption may not hold even approximately. When verification is nonignorably missing, the most general model of the distribution of disease state, test result, and verification indicator is overparameterized. Parameters are only partially identified, creating regions of ignorance for maximum likelihood estimators. For studies of a single test, we use Bayesian analysis to implement the most general nonignorable model, a reduced nonignorable model with identifiable parameters, and the MAR model. Simple Gibbs sampling algorithms are derived that enable computation of the posterior distribution of test accuracy parameters. In particular, the posterior distribution is easily obtained for the most general nonignorable model, which makes relatively weak assumptions about the missing data mechanism. For this model, the posterior distribution combines two sources of uncertainty: ignorance in the estimation of partially identified parameters, and imprecision due to finite sampling variability. We compare the three models on data from a study of the accuracy of scintigraphy to diagnose liver disease.