Prevalence-dependent diagnostic accuracy measures.

We study prevalence-dependent diagnostic accuracy measures, specifically, positive and negative predictive values. These measures permit an assessment of the clinical utility of diagnostic tests across populations with different disease prevalences. In many cases, prevalence may not be known with certainty and the evaluation of the diagnostic tests must account for this uncertainty. A sensitivity analysis may be desired across a prevalence continuum defining low to high-risk populations. For this scenario, simultaneous inference about the predictive values across a range of prevalences is proposed. For situations where a non-point prior distribution on prevalence is specified, we suggest inferences based on averaging the accuracy measures with respect to the prior, leading to simple numerical summaries. The methods are illustrated in a meta-analysis of diagnostic tests for intravascular device-related bloodstream infection, where the prevalence may vary widely both within and across populations.