Personally, I believe that people overuse the fixed-effect model. When you use that model, you are assuming that there is only one effect and it is the same among all studies in your meta-analysis. To do that, you are assuming that all studies were conducted very similarly and we know that occurs rarely. With an I2 of 75%, you have a lot of between-study variance (at least relative to total variance) in effect size and I would consider a fixed-effect model inappropriate to use for the overall meta-analysis or for the follow-up meta-regression.

Dear Philippa,

I totally agree with Gordon. I just want to add that in the case a fixed-effects model would make sense (very little between-study variance), the random-effects model automatically converges into the fixed-effects model. Another argument to use the random-effects model over the fixed-effects model...

Good luck with the meta-analysis!

Henk

I have a different view from Gordon and Henk :-). Once you have gross heterogeneity, the RE weights become more or less equal and a RE meta-regression becomes an unweighted (equal weights) regression. The only problem with fixed effects when heterogeneity is present is that error estimation is poor, but error estimation with RE is only slightly improved and at the cost of less efficiency. The trade off is not worth it if the error estimation with fixed effect models can be fixed. My suggestion is to run a fixed effects meta-regression in Stata and use robust error variances to get correct error estimation. 

I agree with Gordon and Henk but Suhail has a point as well.

Random effects models do perform better than fixed-effect alternatives in the presence of ANY heterogeneity in simulation settings. See: http://www.ncbi.nlm.nih.gov/pubmed/11252006 and http://www.ncbi.nlm.nih.gov/pubmed/21148194. However, if you have an extremely unusual set-up with a trial with thousands of patients being meta-analysed with one with 50 patients, Suhail is right: the RE model will weight them to be much more similar compared to what the FE model would do (no matter what the heterogeneity estimate is) which is not really acceptable. Hybrid methods are being developed that are trying to use fixed-effect weights rather than the inverse variance weighting that is used in the RE models. On the other hand, I would suspect the 50 subjects trial would have been dropped due to quality concerns at some stage. So sensitivity analyses where you focus on the best quality studies are also important.

However, the size of the meta-analysis DOES affect the heterogeneity estimate, and the smaller a MA the more likely it is you will fail to identify existing heterogeneity. More or less, all heterogeneity estimators perform poorly unless you have >10 studies which is not the norm (see: http://www.ncbi.nlm.nih.gov/pubmed/23922860) So tread carefully with sensitivity analyses.

I have a presentation here which you might find helpful with more information on the pros and cons: https://www.researchgate.net/publication/283574150_Software_and_model_selection_challenges_in_meta-analysis_model_assumptions_heterogeneity_and_individual_patient_data

Presentation Software and model selection challenges in meta-analysis: mo...

Thank you very much for the advice, I really appreciate it.