Summary
Background: Meta-analyses require a thoroughly planned procedure to obtain unbiased overall estimates. From a statistical point of view not only model selection but also model implementation in the software affects the results.
Objectives: The present simulation study investigates the accuracy of different implementations of general and generalized bivariate mixed models in SAS (using proc mixed, proc glimmix and proc nlmixed), Stata (using gllamm, xtmelogit and midas) and R (using reitsma from package mada and glmer from package lme4). Both models incorporate the relationship between sensitivity and specificity – the two outcomes of interest in meta-analyses of diagnostic accuracy studies – utilizing random effects.
Methods: Model performance is compared in nine meta-analytic scenarios reflecting the combination of three sizes for meta-analyses (89, 30 and 10 studies) with three pairs of sensitivity/specificity values (97%/87%; 85%/75%; 90%/93%).
Results: The evaluation of accuracy in terms of bias, standard error and mean squared error reveals that all implementations of the generalized bivariate model calculate sensitivity and specificity estimates with deviations less than two percentage points. proc mixed which together with reitsma implements the general bivariate mixed model proposed by Reitsma rather shows convergence problems. The random effect parameters are in general underestimated.
Conclusions: This study shows that flexibility and simplicity of model specification together with convergence robustness should influence implementation recommendations, as the accuracy in terms of bias was acceptable in all implementations using the generalized approach.
Keywords
Bivariate mixed models - meta-analysis - diagnostic accuracy studies - SAS - Stata
