ipdmetan - estimation_command

[Mark Ebell]

I would like to use ipdmetan to do an IPD meta-analysis for studies with treatment response yes/no as the outcome. The sample code provided shows how to set up the estimation_command for hazard ratios, but not for dichotomous or 2x2 format outcomes. Any help would be much appreciated, and apologies if this is a stupid question! I'm a primary care researcher without access to a biostatistician who works with IPD meta-analysis. David Fisher

[David Fisher]

Hi Mark Ebell ,

ipdmetan is designed primarily for "two-stage" meta-analysis, where the first "stage" is a regression or GLM fitted to each trial in turn; and the second "stage" involves pooling the trial-specific estimation parameters together using standard meta-analysis methods.

So, for example, if we have survival data, typical use would involve Cox regression, e.g.

 . ipdmetan, study(studyname) ... : stcox ...

...and if we have 2x2 data we'd typically use logistic regression, e.g.

 . ipdmetan, study(studyname) ... : logit ...

But, as you will have noticed from reading the documentation, ipdmetan is more flexible than that. If you genuinely wish to perform IPD meta-analysis by directly analysing 2x2 data (rather than fitting regression models or GLMs), you would use

. ipdmetan outcome treat, rr study(studyname) ...

where outcome is the 0/1 outcome indicator variable, and treat is the 0/1 treatment indicator variable. To analyse odds ratios or risk differences (rather than relative risks), replace "rr" in the above with "or" or "rd".

Note that results will be the same as coding:

. gen byte a =  outcome *  treat
. gen byte b = !outcome *  treat
. gen byte c =  outcome * !treat
. gen byte b = !outcome * !treat
. collapse (sum) a b c d, by(studyname)
. metan a b c d, study(studyname)

I hope that helps!
Thanks,
David.

[Mark Ebell]

Thank you, David, much appreciated. I'll dig into this today.

[Mark Ebell]

I am trying to get the results as adjusted odds ratios rather than beta coefficients using this code and your sample dataset, with fail as the dependent variable and trt as the independent:

ipdmetan, study(trialid) re by(region): logit fail trt, or
ipdmetan, study(trialid) re by(region): logistic fail trt

However, both of the above still result in a table of beta coefficients. How do I display the odds ratios, other than just doing the calculation post hoc?

Thanks,

Mark

[David Fisher]

Hi Mark,
The secret is to pass the request for exponentiated betas (that is, odds ratios) to ipdmetan rather than to the estimation_command . That is:

ipdmetan, study(trialid) re by(region) or : logit fail trt
ipdmetan, study(trialid) re by(region) or : logistic fail trt

This makes sense, as we are requesting output from ipdmetan and not from estimation_command .
Thanks,
David

[Mark Ebell]

Thank you! As always, this non-statistician very much appreciates your help.