[R-meta] Does clubSandwich::coef_test() handle crossed random-effects?
Just to distinguish V elements from correlated random effects elements,
when making a multivariate model due to say variable C_ijk (recall I have
'scales' subsuming 'studies' subsuming 'true effects'), then we can assume
Cov(u_{bjk}, u_{cjk}) to be our correlated random effects at the 'study'
level, but could also assume Cov(u_{bk}, u_{ck}) as our correlated random
effect at the 'scale' level, correct?
On Wed, Oct 6, 2021 at 10:04 PM Farzad Keyhan <f.keyhaniha at gmail.com> wrote:
Sure, I think I meant the same thing, I meant the cluster that contains individual effects not higher clusters that contain aggregated effects. Thanks very much, Fred On Wed, Oct 6, 2021 at 9:30 PM James Pustejovsky <jepusto at gmail.com> wrote:
I don't know what "directly and immediately" means. I mean clusters where the sampling errors (or errors of estimation), defined as the difference between the effect size estimate and its target parameter, are correlated. James On Wed, Oct 6, 2021 at 9:26 PM Farzad Keyhan <f.keyhaniha at gmail.com> wrote:
Many thanks, you mean the cluster that "directly and immediately" contains the true and subsequently overlapping observed effects, not the ones higher up in the hierarchy, that is the logic, correct? Fred On Wed, Oct 6, 2021 at 9:21 PM James Pustejovsky <jepusto at gmail.com> wrote:
Hi Fred, The cluster argument in impute_covariance_matrix describes sets of effect sizes that you expect to have correlated sampling errors, which arise if multiple effect sizes are estimated from a common sample (or from partially overlapping samples). So in your case, use cluster = study. James On Wed, Oct 6, 2021 at 9:03 PM Farzad Keyhan <f.keyhaniha at gmail.com> wrote:
Dear James, One quick question, (recall I have 'scales' subsuming 'studies' subsuming 'true effects'). In this case, to set up a V matrix, should I use 'study' as or 'scale' to define the 'cluster' argument in 'impute_covariance_matrix()'? Thanks, Fred On Sun, Oct 3, 2021 at 9:25 PM Farzad Keyhan <f.keyhaniha at gmail.com> wrote:
Dear James,
I explored the issue, there was a re-coding bug. One thing that I
wanted to clarify is that in addition to the 'scale > study' nesting
relationship, the same 'scale' was used to measure different 'outcomes' and
different 'scales' can be used to measure the same 'outcome' across the
studies.
Do you see any potential for crossed random-effects here?
(data attached for clarity)
Fred
dat <- read.csv("
https://raw.githubusercontent.com/ilzl/i/master/j.csv")
study scale yi vi es group outcome time
1 A1 p1 1.680746 0.2081713 1 1 3 0
2 A1 p1 4.122057 0.4806029 2 2 3 0
3 A1 p1 2.600443 0.2838905 3 1 3 1
4 A1 p1 3.457194 0.3836960 4 2 3 1
5 A1 p1 1.546293 0.1998273 5 1 3 2
6 A1 p1 3.071523 0.3352741 6 2 3 2
On Sun, Oct 3, 2021 at 6:59 PM James Pustejovsky <jepusto at gmail.com>
wrote:
On Sun, Oct 3, 2021 at 1:18 PM Farzad Keyhan <f.keyhaniha at gmail.com> wrote:
I see, I'm still exploring to see what has caused the two models in my previous email to give slightly different fits. Still curious though, for 'scale' and 'study' to have been crossed random effects, 'scale' should have varied in each study?
Yes.