Hi all,
I actually feel like I can weigh in here.
While working on a revision for a separate paper on the GVAR, we encountered a reviewer question about clustering with subjects within a group versus the total sample size and how those are somewhat confounded.
We added an additional simulation where we increased total N but varied the within-subgroup sample sizes and found that parameter estimates remained largely the same; to the third decimal place in most cases.
While not GIMME, the algorithm we're using for subgrouping and the method for constructing the adjacency matrices are largely the same and our general conclusions were that the subgrouping is more sensitive to the number of differences between groups in terms of dynamics rather than the number of people within subgroups.
So, I'd echo Katie's point that it's likely a reliable result (i.e., the identified subgroups are indeed different) but would be cautious about the generalizability of those results given the relatively small samples.
All of this being dependent on satisfying power requirements and the like.
Best,
Jonathan J. Park