Bivariate meta-analysis of predictive values of diagnostic tests can be an alternative to bivariate meta-analysis of sensitivity and specificity

OBJECTIVE: Meta-analysis of predictive values is usually discouraged becausethese values are directly affected by disease prevalence, but sensitivity andspecificity sometimes show substantial heterogeneity as well. We propose abivariate random-effects logitnormal model for the meta-analysis of the positive predictive value (PPV) and negative predictive value (NPV) of diagnostic tests.STUDY DESIGN AND SETTING: Twenty-three meta-analyses of diagnostic accuracy were reanalyzed. With separate models, we calculated summary estimates of the PPV and NPV and summary estimates of sensitivity and specificity. We compared thesesummary estimates, the goodness of fit of the two models, and the amount ofheterogeneity of both approaches.RESULTS: There were no substantial differences in the goodness of fit or amountof heterogeneity between both models. The median absolute difference between the projected PPV and NPV from the summary estimates of sensitivity and specificityand the summary estimates of PPV and NPV was 1% point (interquartile range, 0-2% points).CONCLUSION: A model for the meta-analysis of predictive values fitted the datafrom a range of systematic reviews equally well as meta-analysis of sensitivityand specificity. The choice for either model could be guided by considerations ofthe design used in the primary studies and sources of heterogeneity.