Editors: Julian PT Higgins, Jonathan J Deeks and Douglas G Altman on behalf of the Cochrane Statistical Methods Group
Key points
When missing data prevent a study from being included in a meta-analysis (and attempts to obtain the data from the original investigators have been unsuccessful), any strategies for imputing them should be described and assessed in sensitivity analyses.
Non-standard designs, such as cluster-randomized trials and cross-over trials, should be analysed using methods appropriate to the design. Even if study authors fail to account for correlations among outcome data, approximate methods can often be applied by review authors.
To include a study with more than two intervention groups in a meta-analysis, the recommended approach is usually to combine relevant groups to create a single pair-wise comparison.
Indirect comparisons of interventions may be misleading, but methods are available that exploit randomization, including extensions into ‘multiple-treatments meta-analysis’.
To reduce misleading conclusions resulting from multiple statistical analyses, review authors should state in the protocol which analyses they will perform, keep the number of these to a minimum, and interpret statistically significant findings in the context of how many analyses were undertaken.
Bayesian approaches and hierarchical (or multilevel) models allow more complex meta-analyses to be performed, and can offer some technical and interpretative advantages over the standard methods implemented in RevMan.
Studies with no events contribute no information about the risk ratio or odds ratio. For rare events, the Peto method has been observed to be less biased and more powerful than other methods.
16.2 Intention-to-treat issues
16.3 Cluster-randomized trials
16.5 Studies with more than two intervention groups
16.6 Indirect comparisons and multiple-treatments meta-analysis
16.7 Multiplicity and the play of chance
16.8 Bayesian and hierarchical approaches to meta-analysis
16.9 Rare events (including zero frequencies)