[R-meta] setting certain covariances to 0 for the random LHS term
Viechtbauer, Wolfgang (SP)
wo||g@ng@v|echtb@uer @end|ng |rom m@@@tr|chtun|ver@|ty@n|
Wed Sep 4 23:14:14 CEST 2019
I don't understand the specifics of your particular case, but yes, you can set covariances (or rather, correlations) to zero. Argument 'rho' and 'phi' are used for that. I assume you are using struct="UN" -- or rather, struct=c("UN","UN") to be precise -- for those two random effects terms. Take a look at:
and find the part that starts with: "For struct="UN", the values specified under 'rho' should be given in column-wise order." Setting an element of that vector to NA means that the corresponding correlation will be estimated. Setting it to 0 sets it to ... 0!
To illustrate, using a very simple case where there are only two levels for the 'inner' term (so there really is only one correlation):
dat <- dat.berkey1998
V <- lapply(split(dat[,c("v1i", "v2i")], dat$trial), as.matrix)
V <- bldiag(V)
res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat)
# So, if I set rho = NA, then the correlation is estimated (as above):
res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat, rho=c(NA))
# But I can set the correlation also to 0:
res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat, rho=c(0))
# which is actually identical here to:
res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="DIAG", data=dat)
Now if you have more than one correlation, then rho will be a vector. And you can pick and choose which correlations to constrain to 0 and which ones will be estimated.
Argument 'phi' works the same way -- just for the second random effects term.
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-project.org] On Behalf Of Gram, Gil (IITA)
Sent: Thursday, 29 August, 2019 9:57
To: r-sig-meta-analysis using r-project.org
Subject: [R-meta] setting certain covariances to 0 for the random LHS term
Dear Wolfgang and others,
In my current model I have the following random structures: (treatment | site) and (treatment | site.time), i.e. random variance across site and across time-within-site for the different treatments.
Treatments here are defined as (1) control, (2) mineral input MR, (3) organic input OR, and (4) combined mineral and organic input ORMR.
Now I extend the number of treatments by the organic input sources, i.e. control, MR, ORone, ORtwo, ORthree, ORMRone, ORMRtwo, ORMRthree etc.
I wish to know the variances for each of these extended treatments, but I don’t want to see covariances between combinations of organic input sources, only with Control (and MR). Is it possible to let the model know that, by setting these to 0 somehow?
PhD researcher | Natural resource management/CCAFS
International Institute of Tropical Agriculture (IITA)
East African Hub/Kampala/Country office
Address: IITA-Uganda Plot 15B, Naguru East Road. P.O Box 7878, Kampala
Mobile: +256 755 315236 | Belgium Mobile: +32 484 981200
More information about the R-sig-meta-analysis