Title: A pseudo-likelihood approach for multivariate meta-analysis of test accuracy studies with multiple thresholds
Authors: Duc Khanh To - University of Padova (Italy) [presenting]
Annamaria Guolo - University of Padova (Italy)
Abstract: In meta-analysis of test accuracy studies, the multivariate approach is an effective technique to synthesize results when each study reports sensitivities and specificities at different thresholds. The normal multivariate mixed-effects model proposed in literature as an extension of the well-known bivariate model for the one threshold case, despite interesting features, suffers from some drawbacks. They include the requirement of an estimate of within-study correlations between sensitivities (and specificities) at different thresholds and convergence issues. In order to overcome such drawbacks, we propose a pseudo-likelihood approach under a working independence assumption between sensitivities (and specificities) at different thresholds in the same study. The approach does not require the within-study correlations to be known or estimated and convergence issues very rarely occur. In addition, its implementation is straightforward. The problem of different set of thresholds per study is taken into account by assuming that the thresholds in each study are missing completely at random and adopting an available case. Several simulation studies show that the proposed method performs satisfactorily and improves on the corresponding results from the normal multivariate mixed-effects model. In order to illustrate the applicability of the pseudo-likelihood approach, some published meta-analyses of diagnostic test accuracy have been analyzed.