Bayesian meta-analysis of diagnostic tests allowing for imperfect reference standards.

There is an increasing interest in meta-analyses of rapid diagnostic tests (RDTs) for infectious diseases. To avoid spectrum bias, these meta-analyses should focus on phase IV studies performed in the target population. For many infectious diseases, these target populations attend primary health care centers in resource-constrained settings where it is difficult to perform gold standard diagnostic tests. As a consequence, phase IV diagnostic studies often use imperfect reference standards, which may result in biased meta-analyses of the diagnostic accuracy of novel RDTs. We extend the standard bivariate model for the meta-analysis of diagnostic studies to correct for differing and imperfect reference standards in the primary studies and to accommodate data from studies that try to overcome the absence of a true gold standard through the use of latent class analysis. Using Bayesian methods, improved estimates of sensitivity and specificity are possible, especially when prior information is available on the diagnostic accuracy of the reference test. In this analysis, the deviance information criterion can be used to detect conflicts between the prior information and observed data. When applying the model to a dataset of the diagnostic accuracy of an RDT for visceral leishmaniasis, the standard meta-analytic methods appeared to underestimate the specificity of the RDT.