Bayesian interval estimation for predictive values from continuous diagnostic tests in case-control studies

Diagnostic tests based on continuous biomarkers often depend on an estimated cutoff point that dichotomizes the biomarkers as diseased and non-diseased. Positive predictive and negative predictive values (PPV and NPV) can be used to evaluate the diagnostic accuracy of such tests. PPV and NPV are functions of test accuracy and disease prevalence. Previous studies providing joint estimation of the predictive values assume a fixed prevalence. However, prevalence is often unknown. Using a Bayesian prior allows for uncertainty in estimating the population prevalence. We use the Alzheimer’s Disease Neuroimaging Initiative (ADNI) data set to compare our new approach to an existing method and to demonstrate how different prevalence estimates affect the interval estimation. Finally, simulation results indicate that increased levels of uncertainty result in wider intervals that are more robust to deviations from the estimated prevalence.