Hierarchical meta-regression: a unified approach to model multiplicity of bias in combining randomized and non-randomized evidence in meta-analysis
Final Report Abstract
Meta-analysis methods allow researchers to combine results from multiple pieces of evidence into a single analysis. These techniques are used to extend the scope of a single study by combining results from several ones. In particular, a meta-analysis of randomized controlled trials (RCTs), addressing the same primary research question, is considered the gold standard in clinical evidence. However, there are important reasons to include observational studies (OS) in a meta-analysis. In particular, RCT’s results might not be representative of real-world populations. Therefore, by combining results of RCTs with OS we can gain new insights from RCTs’ results that cannot be seen using only a meta-analysis of RCTs. Moreover, while a clinical question may be simple, (e.g., what is the effect of an intervention in a population of interest?), the structure of evidence available to answer it might be limited to a few published results coming from different study types. Consequently, we need meta-analysis methods to combine disparate pieces of evidence to gain information in order to answer such questions. The aim of this project was to develop a new meta-analysis approach: the Hierarchical Meta-Regression (HMR), which can account for a multiplicity of biases in combining RCTs and OS in meta-analysis. This method can be used to combine results of different study types with the same outcome variable and similar interventions. The study types considered were RCTs and OS, covering studies with non-randomized control groups and studies without a control group (e.g. a register or cohort study). In addition, we considered the case of combining aggregated data with individual participant data. Two forms of biases were modeled within the HMR framework: the internal validity bias and the external validity bias. The internal validity bias threatens study results in a way that the observed results are a biased estimate of the effect of an intervention. The external validity bias limits the degree to which the study results are applicable to other populations and settings. We prepared several case studies in meta-analysis, which provided examples of controversial results. We found that commonly used meta-analysis procedures were not robust against internal and external validity bias. Consequently, researchers arrived to misleading conclusions. The application of the HMR resolved these issues by incorporating a bias model into the metaanalysis. The results of applying the HMR were validated by several simulation experiments. In addition, during this project, a new free available R package was developed: jarbes, which stands for “Just a rather Bayesian evidence synthesis”. This package allows the application of the HMR approach and contains the data of the case studies investigated in this project. An important conclusion of this project is that ignoring internal and external validity bias in metaanalysis is a form of model misspecification that conduces to results that could be misleading. The HMR approach is a possible solution to this important problem in meta-analysis.
Publications
-
(2017). Two Examples of Bayesian Evidence Synthesis with the Hierarchical Meta-Regression Approach. Chap. 9, 189-206. Bayesian Inference, ed. Tejedor, Javier Prieto. InTech.
Verde, P. E.
-
(2018). bamdit: An R Package for Bayesian Meta-Analysis of Diagnostic Test Data. Journal of Statistical Software, 86 (10), 1 - 32
Verde, P. E.
-
(2018). The hierarchical meta-regression approach and learning from clinical evidence. Biometrical Journal. 1 - 23
Verde P. E.
