[R-meta] Metafor: meta regression using rma function for proportion with categorical and continuous variable using PFT transformation - R-SIG-meta-analysis

Tue, Oct 1, 2024 12:12 AM #

You need to give the 'at' argument to emmprep(), which is a wrapper around emmeans::qdrg(), which in turn is a wrapper around emmeans::ref_grid(). Here is an example using identical models fitted with rma() and lm() to show that this yields identical results:

library(metafor)
library(emmeans)

dat <- dat.debruin2009
dat <- escalc(measure="PLO", xi=xi, ni=ni, data=dat)

res1 <- rma(yi, vi=0, mods = ~ scq + ethnicity, data=dat, test="knha")
grd <- emmprep(res1, at=list(scq=12))
emmeans(grd, pairwise ~ scq + ethnicity, type="response", infer=c(TRUE,TRUE))

res2 <- lm(yi~ scq + ethnicity, data=dat)
emmeans(res2, pairwise ~ scq + ethnicity, tran="logit", type="response", infer=c(TRUE,TRUE), at=list(scq=12))

Best,
Wolfgang

-----Original Message-----
From: Reza Norouzian <rnorouzian at gmail.com>
Sent: Tuesday, October 1, 2024 03:07
To: R Special Interest Group for Meta-Analysis <r-sig-meta-analysis at r-
project.org>
Cc: Viechtbauer, Wolfgang (NP) <wolfgang.viechtbauer at maastrichtuniversity.nl>
Subject: Re: [R-meta] Metafor: meta regression using rma function for proportion
with categorical and continuous variable using PFT transformation

Wolfgang,

I typically use emmeans for post model comparisons of this kind. But in applying
this to your example, I noticed the "at=" argument in emmeans(), to hold say
_scq_ at 12, is not coming through from emmprep() --at least for me-- at all. If
this is true for you all as well, then it could possibly be some bug lurking
around somewhere in emmprep() or qdrg().

Reza

dat <- dat.debruin2009
dat <- escalc(measure="PLO", xi=xi, ni=ni, data=dat)

res <- rma(yi, vi, mods = ~ scq + ethnicity, data=dat)
grd <- emmprep(res)

emmeans(grd, pairwise~scq + ethnicity, type="response", infer=c(TRUE,TRUE),
at=list(scq=12))

Compare to:

res2 <- lm(yi~ scq + ethnicity, data=dat)

emmeans(res2, pairwise~scq + ethnicity, tran= "logit",
type="response",  infer=c(TRUE,TRUE),
        at = list(scq = 12))

On Mon, Sep 30, 2024 at 2:11?AM Viechtbauer, Wolfgang (NP) via R-sig-meta-
analysis <mailto:r-sig-meta-analysis at r-project.org> wrote:
Dear Daidai,

Using the FT transformation for a meta-analysis is actually quite problematic:

Schwarzer, G., Chemaitelly, H., Abu-Raddad, L. J., & R?cker, G. (2019).
Seriously misleading results using inverse of Freeman-Tukey double arcsine
transformation in meta-analysis of single proportions. Research Synthesis
Methods, 10(3), 476?483. https://doi.org/10.1002/jrsm.1348

R?ver, C., & Friede, T. (2022). Double arcsine transform not appropriate for
meta?analysis. Research Synthesis Methods, 13(5), 645?648.
https://doi.org/10.1002/jrsm.1591

So I would generally advise against its use. Good alternatives are the standard
arcsine or logit transformations.

The appropriate back-transformations are transf.iarcsin() and transf.ilogit()
(the latter is really just plogis()). But you cannot back-transform the
coefficients with these transformations, since these are non-linear
transformations (so the influence of one moderator depends on the values of the
other moderators). You can however back-transform predicted values based on the
model. Consider this example:

library(metafor)

dat <- dat.debruin2009
dat <- escalc(measure="PLO", xi=xi, ni=ni, data=dat)

res <- rma(yi, vi, mods = ~ scq + ethnicity, data=dat)
res

# compute predicted values for scq=11 and scq=12 for ethnicity=caucasian
out <- predict(res, newmods=cbind(c(11,12),0), transf=transf.ilogit)
out
out$pred[2] - out$pred[1]

# compute predicted values for scq=11 and scq=12 for ethnicity=other
out <- predict(res, newmods=cbind(c(11,12),1), transf=transf.ilogit)
out
out$pred[2] - out$pred[1]

Notice that the difference between predicted values that differ by one unit on
scq depends on ethnicity.

For logit-transformed proportions, you can however exponentiate the
coefficients, which then correspond to odds ratios:

round(exp(coef(summary(res)))[c(1,5,6)], digits=2)

So, a one-unit increase in scq corresponds to a 1.05 times increase in the odds
(or the odds are 5% higher when scq is one unit higher) and this holds true
whether ethnicity=caucasian or ethnicity=other.

Best,
Wolfgang