[R-meta] Specifying random effect structure (nested and cross-classified)
Dijk, C.N. van (Chantal)
C@N@v@nD|jk @end|ng |rom |et@ru@n|
Tue Nov 26 11:08:07 CET 2019
I have a question about my random effects structure. I have a dataset consisting of 198 effect sizes (Hedges' g) drawn from 38 experimental groups, which were drawn from 27 studies (for an example of the structure, see: https://www.scribd.com/document/436956335/Example-Data-Structure-20191114?). There are multiple comparisons within studies due to multiple outcomes per experimental group. This situation is similar to the one described in an earlier post: https://stat.ethz.ch/pipermail/r-sig-meta-analysis/2019-July/001624.html. What I understood from Wolfgang's answer is that I can specify my random effect structure as:
res1 <- rma.mv(yi, vi, random = ~ 1 | study/group/effectid, data=dat)
However, my data structure is more complex in the sense that many papers report on results from different tasks. I cannot simply add an additional level, because experimental group is not nested in task (or vice versa): effect sizes from multiple experimental groups can be reported for the same task and effect sizes from multiple tasks can be reported for the same experimental group. Therefore it seems to me I need to specify a cross-classified random effects structure (in addition to the nested structure). In lme4 (lmer) I could specify the random effects as:
(1|study/group/effectid) + (1|effectid:(group:data_collection)) + (1|group:data_collection)
My questions are:
- Does it make sense to specify the dependencies in the data like this?
- If it does, how can I specify this structure for rma.mv?
- Is there a better way to model the dependencies in my dataset??
Many thanks in advance,
Chantal van Dijk | PhD candidate | Centre for Language Studies - Radboud University Nijmegen | Erasmus building (room 8.16) | 024-3616069 | www.ru.nl/2in1project
[[alternative HTML version deleted]]
More information about the R-sig-meta-analysis