Procedures for performing meta-analyses of the accuracy of tools for binary classification

The assessment of accuracy in binary classification tools must take into account two non-independent rates: true positives and false positives. A variety of indices have been proposed. They have been estimated for tests employed for early detection or screening purposes. We summarize and review the main methods proposed for performing a meta-analysis that assesses the accuracy of this type of tools. They are applied to the results from 14 studies that report estimates of the accuracy of the AUDIT. The method of direct aggregation does not allow the use of meta-analytic procedures; the separate estimation of sensitivity and specificity does not acknowledge that they are not independent; the SROC method treats accuracy and threshold as fixed effects and has limitations to deal with the potential role of covariates. The Normal Bivariate (NB) model and the Hierarchical Summary ROC (HSROC) model are statistically rigorous and can deal with the covariates properly. They allowed analyzing the association between the gender composition of the sample and the way the test AUDIT behaves in the example.