Background: To develop appropriate strategies for drawing conclusions from meta-analyses and changing them as necessary when systematic reviews are updated.

Methods: A review of statistical methods for updating meta-analyses. Data: We take examples resulting from a previous survey of Cochrane reviews in which meta-analyses were updated over time. Examples are chosen to illustrate different problems associated with updating reviews, including deciding when there is sufficient evidence in favour, or not in favour, of an intervention, and dealing with apparently changing conclusions.

Results: The default Cochrane 'method' of undertaking a conventional meta-analysis at each update is easily implemented but leaves the reviewer to make subjective judgements about which it is difficult to be explicit. Classical (frequentist) statistical methods that formally incorporate the sequential nature of updated reviews require pre-specification of quantities (such as a clinically worthwhile improvement), which may be inappropriate in a changing environment and on a level that is globally relevant. Bayesian approaches provide a succinct summary of the data that is of use to a variety of readers, and can incorporate updates without pre-specification of quantities. They may be viewed as transparent, explicit and intuitive, but are not yet readily implemented in software.

Conclusion: A Bayesian approach to meta-analyses within Cochrane reviews would allow meta-analyses to be updated without adding complexities to the statistical methods or the subjective interpretation of results. Bayesian meta-analysis can also formally help inform decisions on the need for future studies, and how large they should be. At present, a major limitation is the availability of suitable software.