[R-meta] random effects for both control group and experimental group
Elise van Wonderen
e@v@nwonderen @end|ng |rom uv@@n|
Fri Sep 23 12:40:37 CEST 2022
For some studies in my dataset multiple experimental groups are compared to a single control group, whereas for other studies multiple control groups are compared to a single experimental group. In addition, quite a few studies report multiple outcomes for each experimental-control group combination.
I was wondering whether it makes sense to include random effects for both groups like this (if this does not overparameterize the model):
rma.mv(yi, V, random = list(~ 1 | study/exp_group/effectid, ~1|contr_group), data = data)
Note that I could have swapped contr_group and exp_group here, as contr_group is actually also nested in study (and effectid is also nested in contr_group). However, contr_group is not nested in exp_group or vice versa, which is why I added it as a crossed random effect.
[To clarify: V is a covariance matrix in which the covariance between effect sizes sharing a common reference group is computed according to Gleser & Olkin (2009), and for effect sizes coming from the same groups I will assume a constant correlation and conduct sensitivity analyses.]
Alternatively, I could create a variable "sample" that identifies effect sizes that are expected to have correlated sampling errors because either both the experimental group and the control group are the same, or only one of the two (see data example below). In that case, I might only include a random effect for sample like this: rma.mv(yi, V, random = ~ 1 | study/sample/effectid, data = data)
study exp_group contr_group sample effectid
1 1 1 1 1
1 1 1 1 2
1 2 2 2 3
2 1 1 3 4
2 1 2 3 5
3 1 1 4 6
3 2 1 4 7
I'd be very happy to hear your thoughts!
Elise van Wonderen, MSc
Amsterdam Center for Language and Communication
University of Amsterdam
More information about the R-sig-meta-analysis