ipdmetan questions

[James Rooney]

Hi all,

I have survival data from a number of different disease registers that I am analyzing using stpm2. I wish to perform an ipd meta-analysis of a particular variable - however despite reading the help files and some online lectures I found, I remain a bit confused about how ipdmetan deals with complex multivariate models. These are perhaps best answered with specific questions around a hypothetical model:

Suppose this is my hypothetical model:

* Random Effects
xi: ipdmetan, study(country) hr forest ///
    re:  stpm2 i.binary i.categorical1 i.categorical2 c.age c.other, ///
            df(3) sc(h)

I have three questions:

1. This produces a forest plot - but which level of which variable is it plotting and can I control this ? (I assume it is 1.binary but I am not 100% sure)
2. Instead of a binary, can I do an ipdmetan of the nth level of a categorical variable (or I should use indicator variables to do this) ?
3. For a complex model - can I extract the coefficients from the other covariates after the model has run? (i.e. the ones that are not in the forest plot). And are these also meta-analysed ?

I hope these questions make sense as I am new to the forum and this command and it is my first time to try to do a meta-analysis.

[David Fisher]

Hi James,

Rather than consider a hypothetical model, it's easier to use a publically-available dataset that everyone can load and run commands on. Here is an example using the dataset provided with the ipdmetan package (incidentally, I'm assuming you have the latest version of ipdmetan from SSC, that is version 1.07, 29jun2015), sticking as close as possible to your hypothetical code:

use "http://fmwww.bc.edu/repec/bocode/i/ipdmetan_example", clear
stset tcens, f(fail)
ipdmetan, study(trialid) hr nograph re : stpm2 i.trt i.stage sex age, df(3) sc(h)

(Note that since Stata version 11, you don't need to prefix the command with "xi").

To answer your questions:

1. You can see which variable it is pooling in the output window. Above the table of figures it should say "Meta-analysis pooling of main (treatment) effect estimate xb:1.trt using Random-effects; DerSimonian-Laird estimator". This tells you that it is the stpm2 coefficient xb:1.trt that is being pooled. In other words, you were correct in your assumption. If ipdmetan is not pooling the coefficient you expected, you can specify it explicitly using the poolvar() option.

2. You can certainly do that -- again, by using the poolvar() option (and again, since Stata version 11 there is no need to manually create indicator variables). For example, to pool the third level of stage in the example dataset:

ipdmetan, study(trialid) hr nograph re poolvar(3.stage) : stpm2 i.trt i.stage sex age, df(3) sc(h)

3. No -- ipdmetan is a "two-stage" meta-analysis command, and simply loops over the studies (or countries, etc.), fits the specified model to each study's (or country's) data, saves the specified coefficients, and pools them. No other coefficients are saved. If you wish to simultaneously meta-analyse several coefficients at once, mvmeta (from SSC) or a true "one-stage" model might be the way to go, since repeated calls to ipdmetan will not take account of covariances between coefficients.

Hope that helps,

David.
(author of ipdmetan)

[James Rooney]

Hi David,

I was hoping you would reply - thank you!

Re indicator variables - yes I only recently figured out the post-Stata 11 convention and actually just forgot to remove xi: here. Thanks the answers make sense and I now wonder why I could not figure it out before.

Thank you!
James

[Claudia Suemoto]

Dear David,
Thank you very much for developing ipdmetan. It is very useful!
I need to pool the survival rate from the KM curves across 5 studies using a random-effect model.
Is it possible to know this using the ipdmetan command?
Best,
Claudia

[David Fisher]

Hi Claudia,
As this is a new topic (pooling of survival rates), I've answered your query in a new post entitled "Meta-analysis of survival rates" so that others are more likely to reply, or to be able to find the post in the future.
Thanks,
David.

[Claudia Suemoto]

Thank you very much David. I will look at your answer there.
Best,
Claudia