Possible to combine metan results into one forest plot?

[Talia Sherrard]

Hello, I was hoping someone would be able to advise on the below

I am running analyses in three cohorts, with 4 models for different sets of covariates

I would like to present my results on a forest plot as follows;

Study 1 unadjusted
Study 2 unadjusted
Study 3 unadjusted
Study 1 M1
Study 2 M1
Study 3 M1
Study 1 M2 ...etc

With the pooled effect size and p values available for each model

Is this possible in stata?

Many thanks for any tips or advice!

[David Fisher]

Dear Talia,

Can I assume that you have individual-level data for all three cohorts? If so, then it sounds like the ipdmetan command is what you need. Type ssc describe ipdmetan at the Stata command line (or ssc install ipdmetan to install it).

You would then need to run three meta-analyses, one for each of your four models; and save the results in four new datasets. Then you can open and append the datasets to form a single dataset, and finally produce a forest plot using the forestplot command (which is included with ipdmetan).

There are extensive help files for both ipdmetan and forestplot available once you have installed the package, but feel free to ask further questions here as well. (Note that if you have a more specific question, it's best to provide some example data and/or code that we can work with.)

Hope that helps!

David.

[Talia Sherrard]

Hi David,

Many thanks for your quick reply! I only have individual level data for two of the cohorts unfortunately, so was trying to go about it as follows;

Data frame set out with regression output from each cohort;

Study	model	coefficient	SE	p-value
A	Unadjusted	0	0	0
A	Model 1	0	0	0
A	Model 2	0	0	0
A	Model 3	0	0	0
B	Unadjusted	0	0	0
B	Model 1	0	0	0
B	Model 2	0	0	0
B	Model 3	0	0	0
C	Unadjusted	0	0	0
C	Model 1	0	0	0
C	Model 2	0	0	0
C	Model 3	0	0	0

Running the command metan coeff stderr, by(model)

This produces a basic forest plot by each group (with lots of cosmetic issues which I will fix!), but also pools the results of each model at the bottom- so I think I am looking for a similar command, which does not do the final pooling?

Please could I also check, if you do have individual level data, what are the benefits of using ipdmetan over manually creating a dataset with summary data? Is going about it with the above approach considered inaccurate?

[David Fisher]

Hi Talia,

Thanks for clarifying your data structure. So I assume you have coefficients and SEs from the four models even in the cohorts for which you don't have individudal-level data?

I'll answer your last question first: ipdmetan is simply a time-saving tool for fitting the same model to multiple cohorts within the same dataset, optionally pooling (meta-analysis), and plotting the results. There are no real statistical benefits over doing the same work manually.

ipdmetan can, in fact, handle both individual-level and cohort-level data simultaneously: if the cohort-level data is stored in a separate Stata dataset, then ipdmetan can read it in, and present it alongside the results of fitting a model to the individual-level data. But you already have your cohort-level data prepared, so on this occasion you do not need this option.

If you are happy to use metan, then to answer your main question, all you need are the options nooverall and nosubgroup , which suppress pooling within subgroups and overall, respectively. (There are also other options which suppress other features of the plot.)

I will also mention the alternative commands admetan and forestplot, which I wrote along with ipdmetan. These commands aim to give more flexibility than metan in terms of the layout of the forest plot (whether or not you are interested in pooling). Since you have subgroups of data (that is, the studies) I would recommend taking a look at admetan (as with metan, using the options nooverall and nosubgroup). For more advanced manipulation of the plot, you might read particularly the documentation for the saving() option and for the forestplot command.

Hope that helps,

David.

[Sumon Khan]

Dear All,
I am trying to do a meta-analysis of randomized control trial studies where I have raw data to calculate two separate odds ratios for baseline and end-line. The "metan" command in Stata simply could be used to summarize these results in two different forest plots. However, What I am trying to do is to present baseline and end-lines results in a single forest plot. Does anybody have any suggestion on how I do that?

Thanks in advance!!

Nuruzzaman