Random effects model for meta-analysis of rare events

Question

The Cochrane Handbook states that

incorporation of heterogeneity into an estimate of a treatment effect should be a secondary consideration when attempting to produce estimates of effects from sparse data

This has for example motivated authors of this meta-analysis to use a fixed effects model for their analysis, because their events of interest were rare.

In general, to what extent should this recommendation be followed? If I have data from studies that I have very good reasons to think that they are heterogeneous, but that report very rare events, what type of model should I use?

Answer 1

There is an updated version of 'the Handbook' with a more detailed explanation (https://training.cochrane.org/handbook/current/chapter-10#section-10-4-4-3). What they're saying that the RE models in Revman can be biased and therefore to use a Peto-OR (FE model) for rare events (<1%). Others (https://pubmed.ncbi.nlm.nih.gov/18562399/) have shown that approximate/ asymptotic methods in general are biased and exact methods are more precise for dealing with meta-analysis of rare events.

Regarding FE vs. RE in general, this Q & A might provide some additional insight (Reporting results of fixed effect MA alongside results of random effects MA). In summary, the choice of FE or RE should be an a priori decision (when both are valid options - not particularly in this case).