Network meta-analysis: introduction and an example that compares devices for PFO closure.

Over the last 20 years, the number of published meta-analyses has exploded and, therefore, clinicians have become more familiar with this study design. The results of different trials are statistically combined to produce a more precise estimate [i.e. an estimate with a narrower confidence interval (CI)]. When a meta-analysis is performed after a credible systematic review, the results become more reliable because this precise estimate represents the totality of relevant evidence and can apply to a wider range of patients. Clinicians have become familiar with these traditional metaanalyses that compare two interventions. However, what if we have three or more potentially effective interventions? A newer type of meta-analysis called network meta-analysis (NMA; also called multiple treatment comparison meta-analysis) addresses such comparisons. We will explain the NMA approach and evaluate a recent example published in this issue of the European Heart Journal. When multiple experimental treatments have not been compared in head-to-head trials or only compared in just a few head-to-head trials—which is a very common scenario—one can infer the effect of intervention A on intervention B from comparing the effects of A and B on a common comparator C (usually placebo). Therefore, we can indirectly compare A and B even if they were not compared in a trial. If direct evidence and indirect evidence exist, they can both be combined to provide a final estimate; this is called NMA. This is demonstrated in Figure 1. The output of NMA includes estimates of effect sizes for all possible pairwise comparisons (e.g. relative risks of A vs. B, A vs. C, C vs. B, etc.) and can include a probability ranking (e.g. probability of intervention A being most effective). The analysis can be done using the most frequent (classic) statistical methods (i.e. the wellknown methods of hypothesis testing and CIs) or using Bayesian statistics (i.e. updating a prior probability with newer evidence to produce a posterior probability). Commonly Bayesian meta-analysis uses a vague (i.e. non-informative) prior, which leads to the results of both methods being very similar.