Causal Meta-Analysis : Methodology and Applications

Abstract

Meta-analysis is a statistical method to summarize research data from multiple studies in a quantitative manner. This dissertation addresses a number of methodological topics in causal meta-analysis and reports the development and validation of meta-analysis software. In the first (methodological) part, we explore how experienced or inexperienced authors of meta-analyses are and show that they are generally not as experienced as one might expect. Next, we discuss the terminology related to biases that occur due to selective dissemination of evidence. In the proposed epistemological framework, publication bias is a pars pro toto and dissemination bias is a more appropriate term to describe the full spectrum of biases. The chapter on the usefulness of graphical assessments shows that forest plots and residual histograms may be best suited to the exploration of heterogeneity in meta-analysis data sets. None of the plots detected (dissemination) selectivity adequately, including the frequently used funnel plots. The last two chapters of the methodology part introduce software (MIX) for causal meta-analysis and report its development and validation, and a comparison with existing meta-analysis software. The applications part contains two meta-analyses in which the previously discussed methods are applied and the MIX software is used. The first meta-analysis re-analyzes data from a recent, widely publicized meta-analysis of 42 clinical trials which concluded that rosiglitazone was associated with an approximately 43% increased risk for myocardial infarction and an approximately 64% increased risk for cardiovascular death. We show that alternative meta-analytic approaches consistently yield results that are not statistically significant and conclude that the risk for myocardial infarction and death from cardiovascular disease for diabetic patients taking rosiglitazone is uncertain: neither increased nor decreased risk is established. The other meta-analysis explores the effects of neuromuscular electrical stimulation on muscle strength, showing that active exercises are preferred over electrical stimulation but that electrical stimulation can have an added benefit in immobilized patients. The dissertation concludes with a practical methodological paper on modern methods of causal meta-analysis, providing a step-by-step guide on how to apply a wide range of meta-analytical methods to causal research data. It describes how a meta-analysis preferably starts with a data exploration, which is then followed by a synthesis and an evaluation of the applied meta-analytical methods.