[R-meta] Inverse weighting after estimation of VCOV
Hi David, I don't entirely understand the models that you're looking at, so clarifying the following would help in getting good feedback: * What is the variable `shared_variance` used in the vcalc call? * What is the variable `number` used in the random effects argument of rma.mv? * How are these variables related? Additionally, it would be good to check that the vcov matrix created by vcalc() is as you intend it to be. Could you pull out the blocks of this matrix for a few studies and just verify that they give you covariance matrices with a correlation of 0.7? I mean something like: vcov_study_k <- V_mat[i:j, i:j] cov2cor(vcov_study_k) where the indices i:j are the rows in your data corresponding to a given study k. James On Fri, May 24, 2024 at 10:00?AM David Pedrosa via R-sig-meta-analysis <
r-sig-meta-analysis at r-project.org> wrote:
Dear all,
I have a basic question about the output of my (gu)estimation of the
variance-covariance matrix. I have extracted results from very
heterogeneous studies with OR as effect size (sample sizes between 20
and 300,000). Since some of the results come from the same study, I
decided to try to use the VCOV as an input and estimated values
according to the following formula
V_mat <- vcalc(vi=vi, cluster=shared_variance, data=df_complete, rho=.7)
res_meta <- rma.mv(yi, vi, V=V_mat,
random = ~ 1 | number, mods = ~ hospitalbeds +
ltcbeds, verbose=TRUE, data=df_complete)
Interestingly, in this case the weighting is reversed, so that most of
the weight is given to studies with the smallest sample size; something
that does not happen when using this formula:
res_meta <- rma(yi, vi,
random = ~ 1 | number, mods = ~ hospitalbeds +
ltcbeds, verbose=TRUE, data=df_complete)
I have tried to understand what is going on, but I am at kind of lost.
Could someone please give me some advice?
Thanks in advance,
David
_______________________________________________ R-sig-meta-analysis mailing list @ R-sig-meta-analysis at r-project.org To manage your subscription to this mailing list, go to: https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis