Systematic review of methods used in meta-analyses where a primary outcome is an adverse or unintended event

Abstract

Background: Adverse consequences of medical interventions are a source of concern, but clinical trials may lack power to detect elevated rates of such events, while observational studies have inherent limitations. Meta-analysis allows the combination of individual studies, which can increase power and provide stronger evidence relating to adverse events. However, meta-analysis of adverse events has associated methodological challenges. The aim of this study was to systematically identify and review the methodology used in meta-analyses where a primary outcome is an adverse or unintended event, following a therapeutic intervention.

Methods: Using a collection of reviews identified previously, 166 references including a meta-analysis were selected for review. At least one of the primary outcomes in each review was an adverse or unintended event. The nature of the intervention, source of funding, number of individual meta-analyses performed, number of primary studies included in the review, and use of meta-analytic methods were all recorded. Specific areas of interest relating to the methods used included the choice of outcome metric, methods of dealing with sparse events, heterogeneity, publication bias and use of individual patient data.

Results: The 166 included reviews were published between 1994 and 2006. Interventions included drugs and surgery among other interventions. Many of the references being reviewed included multiple meta-analyses with 44.6% (74/166) including more than ten. Randomised trials only were included in 42.2% of meta-analyses (70/166), observational studies only in 33.7% (56/166) and a mix of observational studies and trials in 15.7% (26/166). Sparse data, in the form of zero events in one or both arms where the outcome was a count of events, was found in 64 reviews of two-arm studies, of which 41 (64.1%) had zero events in both arms.

Conclusions: Meta-analyses of adverse events data are common and useful in terms of increasing the power to detect an association with an intervention, especially when the events are infrequent. However, with regard to existing meta-analyses, a wide variety of different methods have been employed, often with no evident rationale for using a particular approach. More specifically, the approach to dealing with zero events varies, and guidelines on this issue would be desirable.