Meta-analysis of survival data

[Daisy Fan]

I would like to undertake a meta-analysis of studies that have all used Cox regression for survival analysis. Is there a command for this on Stata?

I found the following paper, which gives one mathematical way of doing it, but I cannot seem to find any papers or reports of how to do this using software. http://onlinelibrary.wiley.com/wol1/...200900168/full

Any help would be gratefully received!

[David Fisher]

Hi Daisy,

Could you be more specific? Do you have IPD or aggregate (summary) data for these studies? If the latter, do you know whether each is adjusted or unadjusted, and do you have reason for thinking that your estimates may be biased in the way that Yuan and Anderson describe? Note that Yuan and Anderson's methods may be of interest in particular circumstances, but also may have disadvantages (e.g increased standard errors, possibility of over-fitting if small number of studies).

If you have aggregate (summary) estimates of logHR and selogHR, the standard Stata meta-analysis command is "metan" (from SSC). This assumes that the estimates are "comparable", i.e. either they are all unadjusted, or adjusted for the same set of parameters, or they are adjusted for differing sets of parameters but that negligible bias is introduced as a result.

If you really want to, I would imagine that Yuan and Anderson's M-A model (I'm not sure about the M-P model) would be relatively easy to fit using "metareg" (also from SSC).

Hope that helps,

David.

[Daisy Fan]

Hi David,

Apologies for the delayed response - I'm new to this forum and didn't realise I'd had a message.

Thank you for your reply. It is aggregate data from around 8-9 studies. Each study contains a range of Cox models, including ones that are unadjusted (or only adjusted for one or two basic demographic variables) and then multiple other models for each study adjusted for various things. The results seem to be similar regardless of what factors are adjusted for, which is good. But I'm wondering if metan for unadjusted models is the best but then perhaps with sensitivity analyses that will allow me to compare the adjusted models (perhaps a mid-level adjusted version and a more adjusted version). Would this sound a reasonable approach?

Kind regards,
Daisy

[Marcos Almeida]

It seems to be a somewhat different situation, but you may consider the aptly named SSC - metainf - so as to "evaluate influence of a single study in meta-analysis estimation".