Cochrane Handbook mentions subgroup analyses are only warranted if there are more than 10 studies (I have 28), and if they are distributed evenly among categories of subgroup. I am wondering if there is any guidance anywhere which is clearer. I.e. what is the uneven distribution ratio cut-off? I have 24 studies in one category and 4 in another, that feels too uneven for me to use, but I would like some rationale behind that please. Or another subgroup has multiple categories, with a split of studies like: 1,1,2,3,5,6,8 - is that ok, only having a single study in some categories?
Any other reference about these subgroup analysis decisions greatly appreciated.
Many thanks!
My personal rule has been a bare minimum of 3 per category. I throw away or combine categories with less. In a subgroup analysis, you're making parametric statistics assumptions. I'd be hard-pressed to calculate a meaningful standard deviation with less than 3 studies. In some large meta-analyses (i.e., >40-50 studies), I have required a minimum of 5 studies per category.
My personal rule has been a bare minimum of 3 per category. I throw away or combine categories with less. In a subgroup analysis, you're making parametric statistics assumptions. I'd be hard-pressed to calculate a meaningful standard deviation with less than 3 studies. In some large meta-analyses (i.e., >40-50 studies), I have required a minimum of 5 studies per category.
Your statistical power will depend on many things so it is difficult to imagine a one-fits-all rule. I agree with Gordon that it can make sense to have a low threshold (as long as you interpret results with sufficient caution). 4-24 seems OK, especially if there is not too much heterogeneity in the smaller group, but the multiple split with several small groups might not provide reliable outcomes.
I added a study in which we decided to not run sub-group analyses if any of the subgroups had too few studies, but using a fairly lenient criterion.
- Badges
- Science topic
- Similar topics
- Psychology
- Emotion