Subgroup differences with metaprop?

[Carla Hope]

Hi,

I am running a metanalysis to look at pooled estimates of attrition using metaprop. I have generated a forest plot as attached. How do I test subgroup differences within this analysis? i.e is there a significant difference between men and women pooled attrition estimate?

Many thanks,
Carla

* Example generated by -dataex-. To install: ssc install dataex
clear
input str20 Study float Year int size byte attrition int completed float(sex totalbysex _ES _seES _LCI _UCI _WT)
"Aufses"     1998   88 11   52 0   63        .125 .035254754   .07125274  .21011756  1.951692
"Bergen"     1998  132 11   92 0  103   .08333334  .02405626   .04716797  .14306453  3.186563
"Brown"      2014   85  9   40 0   49   .10588235  .03337334   .05671241   .1891352 2.1106992
"Carter"     2018   88 11   52 0   63        .125 .035254754   .07125274  .21011756  1.951692
"Dodson"     2004  120 11   76 0   87   .09166667  .02634133   .05195718   .1567085 2.8715415
"Gifford"    2014  371 48  128 0  176   .12938005 .017424528    .0989906   .1673659 4.3181467
"Nadeem"     2014  106 29   55 0   84    .2735849  .04329977    .1977593  .36524725   1.42592
"Sullivan"   2013 2033 81 1236 0 1317    .0398426 .004337868  .032172184  .04924871  6.828648
"Symer"      2018  792 84  421 0  505    .1060606  .01094129   .08648507  .12943918  5.662387
"Yaghoubian" 2012  348 29  191 0  220   .08333334 .014815813   .05864695  .11711818  4.844076
"Yeo"        2017  836 90  438 0  528    .1076555 .010719666   .08841137  .13048883  5.708778
"Yeo"        2010 6303 50 2641 0 2691   .00793273 .001117399  .006022631  .01044226  7.082663
"Aufses"     1998   88  8   17 1   25    .0909091  .03064545   .04678642   .1692539 2.3720863
"Bergen"     1998  132  7   22 1   29    .0530303 .019504873   .02592242   .1054179  3.928538
"Brown"      2014   85  7   29 1   36   .08235294 .029817274   .04046373  .16035984 2.4594166
"Carter"     2018   88  8   17 1   25    .0909091  .03064545   .04678642   .1692539 2.3720863
"Dodson"     2004  120  9   24 1   33        .075  .02404423   .03995712  .13640918 3.1883204
"Gifford"    2014  371 39   73 1  112    .1051213  .01592357    .0778562     .14048  4.616435
"Nadeem"     2014  106 12   10 1   22   .11320755 .030774835   .06595682   .1875127 2.3587935
"Sullivan"   2013 2033 50  666 1  716  .024594195 .003435107  .018704975 .032276634  6.927888
"Symer"      2018  792 69  218 1  287    .0871212  .01002088  .069419935  .10880835  5.853562
"Yaghoubian" 2012  348 26  102 1  128   .07471264 .014094372   .05149312  .10721887   4.99494
"Yeo"        2017  836 74  234 1  308   .08851675   .0098239    .0710965  .10970127  5.893878
"Yeo"        2010 6303 27 1241 1 1268 .0042836745 .000822626 .0029457496 .006225475  7.091249
end
label values sex sex
label def sex 0 "Male", modify
label def sex 1 "Female", modify

------------------ copy up to and including the previous line ------------------

Listed 24 out of 24 observations

[David Fisher]

Hi Carla,


To answer your question as asked:

If you have performed a fixed-effect analysis, I think the "heterogeneity between groups" is equivalent to a test for subgroup effect difference; so p=0.061.

Alternatively, you could perform a meta-regression of effect size on sex using the package metareg (available from SSC). Note that metareg performs random-effects analysis by default, so you may need to use the "z" option (I think?) to get it to agree with your forest plot. Also, you will probably need to feed metareg the calculated effect size variables _ES and _seES generated by metaprop, rather than the raw count data.


However...

Looking at your forest plot, I notice that you have the same studies in both groups, suggesting that you have male and female results for each of your studies. In that case, I strongly recommend reading the following paper (disclaimer, I am first author!!): https://www.bmj.com/content/356/bmj.j573 and consider forming "covariate (i.e. sex) interactions" for each of your studies, and pooling those. Since you have proportion data, I guess your interactions could simply be represented by relative risks or risk differences, which may simplify matters -- this particular circumstance is not one I have encountered before!

If you wish to go down this path, please feel free to ask me additional questions, and I will do my best to help.


Best wishes,

David.

[Hunduma Dinsa]

dear all, why metaprop in stata can not estimate the I2 and P value during subgroup analysis in which the number of studies are 2 or 3. but it estmates for 4 studies or more per group?

[Hunduma Dinsa]

is it because of the version I used or not? Used stata version 14

[David Fisher]

Dear Hunduma Dinsa , could you post an example, so that we can see what you mean? If you look at the first post in this thread, you can see a 2-subgroup analysis with metaprop which clearly shows I2 and P values throughout.

Edit: Sorry, I realised I mis-read your question -- you were asking about subgroups with small numbers of studies, rather than small numbers of subgroups.

So, to answer your actual question: I'm not sure to be honest. Certainly your version of Stata should not be a concern.

My suggestion is to install the latest version of metan from SSC, and to use the proportion option. This will do the same job as metaprop, but will give you the I2 and p-values you require.

In other words, if your existing command is:

metaprop events total, by(subgroup) ...

the new command would be:

metan events total, proportion by(subgroup) ...

Hope that helps.
Best wishes,
David.