A double SIMEX approach for bivariate random-effects meta-analysis of diagnostic accuracy studies

Annamaria Guolo

Bivariate random-effects models represent a widely accepted and recommended approach for meta-analysis of test accuracy studies. Standard likelihood methods routinely used for inference are prone to several drawbacks. Small sample size can give rise to unreliable inferential conclusions and convergence issues make the approach unappealing. This paper suggests a different methodology to address such difficulties. A SIMEX methodology is proposed. The method is a simulation-based technique originally developed as a correction strategy within the measurement error literature. It suits the meta-analysis framework as the diagnostic accuracy measures provided by each study are prone to measurement error. SIMEX can be straightforwardly adapted to cover different measurement error structures and to deal with covariates. The effortless implementation with standard software is an interesting feature of the method. Extensive simulation studies highlight the improvement provided by SIMEX over likelihood approach in terms of empirical coverage probabilities of confidence intervals under different scenarios, independently of the sample size and the values of the correlation between sensitivity and specificity. A remarkable amelioration is obtained in case of deviations from the normality assumption for the random-effects distribution. From a computational point of view, the application of SIMEX is shown to be neither involved nor subject to the convergence issues affecting likelihood-based alternatives. Application of the method to a diagnostic review of the performance of transesophageal echocardiography for assessing ascending aorta atherosclerosis enables overcoming limitations of the likelihood procedure. The SIMEX methodology represents an interesting alternative to likelihood-based procedures for inference in meta-analysis of diagnostic accuracy studies. The approach can provide more accurate inferential conclusions, while avoiding convergence failure and numerical instabilities. The application of the method in the R programming language is possible through the code which is made available and illustrated using the real data example.