Performance measures of the bivariate random effects model for meta-analyses of diagnostic accuracy

The bivariate random effects model has been advocated for the meta-analysis of diagnostic accuracy despite scarce information regarding its statistical performance for non-comparative categorical outcomes. Four staggered simulation experiments using a full-factorial design were conducted to assess such performance over a wide range of scenarios. The number of studies, the number of individuals per study, diagnostic accuracy values, heterogeneity, correlation, and disease prevalence were evaluated as factors. Univariate and bivariate random effects were estimated using NLMIXED with trust region optimization. Bias, accuracy, and coverage probability were evaluated as performance metrics among 1000 replicates in 272 different scenarios. Number of studies, individuals per study, and heterogeneity were the most influential meta-analytic factors affecting most metrics in all parameters for both random effects models. More studies improved all metrics while low heterogeneity benefited fixed and random effects but not the correlation. About twenty studies are required to obtain random effects estimates with good statistical properties in the presence of moderate heterogeneity, while only the univariate model should be used when few studies are summarized. In general, the bivariate model is advantageous for meta-analyses of diagnostic accuracy with complete data only when the correlation is of interest.