[R-meta] multi-model prediction with rma.mv ?

Kelly Gravuer kelly.gravuer at gmail.com
Tue Jan 2 19:42:32 CET 2018

Hello all,
Using the rma.mv() function in metafor, I'm running models with two
random effects and four moderator variables. The format of the model

m1 <- rma.mv(yi ~ mod1 + mod2 + mod3 + mod4, vi, random = list(~ 1 |
rand1, ~1 | rand2), tdist=TRUE, method="ML", data=d)

I would like to do two things: (1) understand the relationship of each
of the four moderator variables to the effect size; (2) predict an
effect size with confidence interval (CI) for some new suites of
values of the moderators.

I have found multi-model inference using glmulti() to be a very useful
approach to my first objective, using the methods described here:
I would now like to use this multi-model suite to make predictions
with CIs.

I realize I can run the model as specified above (including all four
moderators as main effects) and then use predict.rma() to generate
estimated effect sizes + CIs.  I realize I can also calculate an
effect size estimate that is informed by the multiple model fits using
the coefficient estimates in the multi-model coefficient table.
However, I don't know how to get CIs for these effect size estimates
that are informed by the multiple model fits.

I think one way to do this would be to write a predict function that
links rma.mv with glmulti.  I don't feel up to this task myself, but
wanted to inquire whether anyone on this list has written such a
function?  Or alternatively, does anyone have other suggestions for
how to use the information from the multiple model fits to inform the
predicted CIs?

Thanks so much for any thoughts,

More information about the R-sig-meta-analysis mailing list