[R-meta] A potential addition to metafor random-effect structures
ye|eng@y@ng1 @end|ng |rom un@w@edu@@u
Sat Feb 4 04:05:34 CET 2023
If I understand correctly, you are talking about simplifying the configuration of the random effects structure in case your designed model is over-parameterized (this is often the case for small meta-analyses). As far as I know, metafor is quite flexible in imposing constraints on the heterogeneity variance-covariance matrix. Briefly, arguments like sigma2, tau2, rho not only can be estimated from the model but also can be set manually. BTW, Wolfgang created many shorthands of simplified "UN". Those shorthands can meet most of the conditions (at least for my own cases). Unless you want to test whether a specific parameter is "significant" (if this is the case, one can use the likelihood ratio test - under the null hypothesis, the statistics follow the chi-square distribution). If I provided any misleading answers, other experts like Wolfgang, James, and Mike may want to correct me.
From: R-sig-meta-analysis <r-sig-meta-analysis-bounces using r-project.org> on behalf of Reza Norouzian <rnorouzian using gmail.com>
Sent: Saturday, 4 February 2023 5:00
To: R meta <r-sig-meta-analysis using r-project.org>
Subject: [R-meta] A potential addition to metafor random-effect structures
[You don't often get email from rnorouzian using gmail.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
From time to time, I encounter situations where the number of levels
for a categorical variable and/or a combination of such variables are
large enough in each study (i.e., creating high-dimensional joint
effect distributions) that I need to forgo adopting a "UN" structure
associated with those levels in favor of a more restricted structure
Depending on my research goals, however, this strategy may be suboptimal.
Recently, I noticed that the glmmTMB package has added a new
random-effects structure for these cases called the "reduced rank"
structure possible by imposing some restrictions on the matrix of
random-effects to ensure the relevant parameters' identifiability.
I'm not sure how much and/or what kind of assessment(s) of this new
structure currently exists in the methodological literature. But on
its surface, it seems that this might potentially offer some solution
to the problem described above.
Will be glad to hear your thoughts/comments on the potential of this
new structure for multivariate-multilevel meta-regression models
perhaps implemented in rma.mv().
R-sig-meta-analysis mailing list @ R-sig-meta-analysis using r-project.org
To manage your subscription to this mailing list, go to:
[[alternative HTML version deleted]]
More information about the R-sig-meta-analysis