[R-meta] [+externe Mail+] Subgroup analysis best practices

[Röver, Christian]

Dear Tamar,

you were thinking of a case where you have several studies, and for
each you have an overall effect as well as effects within subgroups
(say, effects for male and female patient subgroups). Correct?

As long as you have only two subgroups, the relatively easy solution
would be to compute the contrast for each study and then meta-analyze
those. For example, if you have log odds ratios (log-ORs) for males and
females, you compute the difference between these (which then
corresponds to the log ratio of ORs (log-ROR)) and perform a meta-
analysis of these RORs.

The advantage of this method is that you're actually only making direct
subgroup comparisons "within studies" and then pooling these. Anything
else (i.e., anything involving a direct comparison of the first study's
subgroup A to the second study's subgroup B) would probably be somewhat
controversial, in particular if you are dealing with randomized
controlled studies.

(Note however that you may get discrepancies between this "average
difference" and the "difference of averages" that you would get when
pooling all the male effects and all the female effects separately and
then comparing these two.)

Once you have more than two subgroups, I am not sure there would be an
easy/standard solution. Suppose you have subgroups A, B and C. For each
study, you could pool the effects for B and C, compute the difference
to group A and pool these differences as above. Then do the same for
groups B and C. (From your description, it seems like this was also
what you may have had in mind.) A somewhat implicit assumption would
seem to be that you only expect one of the subgroups to differ from the
remaining two (?). The alternative would be to compute all three
contrasts (A vs. B, A vs. C and B vs. C), and pool these, which in this
case again requires you to perform 3 meta-analyses. Once you have more
than 3 subgroups, the number of possible pairwise comparisons however
increases quickly. Also, you may start running into multiple-testing
problems.

I am wondering whether one might also be able to approach this as some
kind of network-meta-analysis problem (treating a three-subgroup study
like a three-armed trial) ... any ideas anyone?

Cheers,

Christian
 

On Mon, 2024-08-12 at 10:26 -0400, Tamar Novetsky via R-sig-meta-
analysis wrote:
> Hello all,
> 
> I have a dataset with topline and subgroup-level treatment effects,
> and I
> would like to be able to answer questions like "[how much] does this
> subgroup perform better/worse than average?". I know that I can use
> ANOVA
> to test whether there is variation between the groups, but I am
> having a
> hard time figuring out the best way to compare each subgroup to the
> others
> or to the overall effect. I would rather be able to say how a
> subgroup's
> effect compares to the overall effect (or the effect of all people
> not in
> that subgroup), rather than make pairwise comparisons between the
> subgroups. Are there best practices on statistical tests to use for
> this
> purpose?
> 
> Thanks,
> 
> 
> *Tamar Novetsky* *(she/her)*
> Data Scientist I
> Eastern Time Zone
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-meta-analysis mailing list @ R-sig-meta-analysis using r-project.org
> To manage your subscription to this mailing list, go to:
> https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis