Meta-Research on Statistical Methods of Combining Diagnostic Studies

Abstract

The current study has the following aims: systematically investigating the extent to which results of recently published meta-analyses of diagnostic test accuracy could be biased when the authors have applied classical hierarchical models that implicitly assumed normality. Another aim is to compare results with a ready to use Bayesian hierarchical model. Finally, the study aims to assess the impact of its internal validity when classical hierarchical models are used for meta-analyses and results are compared with an Bayesian hierarchical approach. In classical hierarchical models, studies with a high risk for normality assumption may have: non-convergence issues, profile likelihood irregularities, and non-positive-definite random-effect covariance matrix. In Bayesian hierarchical models, the same studies don’t present any difficulties in the fitting or estimation process