[R-meta] Rare dependent variable with correlation among effect sizes

I don't remember our numbering off the top of my head, but yep, that is the model I just described, with fixed study effects and using +-1/2 coding for the treatment random effect. With more than a single treatment group, you will need to think about how to code the latter. But I am off to bed now.

-----Original Message-----
From: Arthur Albuquerque [mailto:arthurcsirio at gmail.com]
Sent: Monday, 06 March, 2023 22:37
To: R Special Interest Group for Meta-Analysis; Viechtbauer, Wolfgang (NP)
Subject: RE: [R-meta] Rare dependent variable with correlation among effect sizes

Following your article "A comparison of seven random-effects models for meta-
analyses that estimate the summary odds ratio?,

would that be "Model 4: a modified version of Simmonds and Higgins model??

glmer(cbind(event,n-event)~factor(study)+factor(treat)+(treat12-1|study),
data=thedata1, family=binomial(link="logit"))

On Mar 6, 2023, 6:33 PM -0300, Viechtbauer, Wolfgang (NP)
<wolfgang.viechtbauer at maastrichtuniversity.nl>, wrote:

When using a logistic model for the analysis, the data structure is changed into
a long / arm-based format. One then adds fixed (or random) effects for studies, a
fixed effect for group, and a random effect for group (to account for
heterogeneity in the treatment effects). This circumvents the issue of a shared
control group. This is in essence the same as what happens in a network meta-
analysis using an arm-based instead of a contrast-based model (in the latter
case, we need to deal with the dependency in three- or more-arm studies, but not
in the arm-based model).

-----Original Message-----
From: Arthur Albuquerque [mailto:arthurcsirio at gmail.com]
Sent: Monday, 06 March, 2023 22:22
To: R Special Interest Group for Meta-Analysis; Viechtbauer, Wolfgang (NP)
Subject: RE: [R-meta] Rare dependent variable with correlation among effect sizes

My effect size of interest is the odds ratio.

A random effect logistic regression with random intercept by study won?t account
for the shared control group within each study.

What other alternative do I have over the sandwich estimator?

On Mar 6, 2023, 6:19 PM -0300, Viechtbauer, Wolfgang (NP)
<wolfgang.viechtbauer at maastrichtuniversity.nl>, wrote:

I don't see the need to use a sandwich estimator, and with 4 studies, this is
unlikely to be all that useful.

-----Original Message-----
From: Arthur Albuquerque [mailto:arthurcsirio at gmail.com]
Sent: Monday, 06 March, 2023 22:01
To: R Special Interest Group for Meta-Analysis; Viechtbauer, Wolfgang (NP)
Subject: RE: [R-meta] Rare dependent variable with correlation among effect sizes

Hi Wolfang, thanks for the quick reply.

About 2), would you fit the model in lme4 then use a sandwich estimator? As you
said, a regular random-effect model in lme4 would be analog to rma.glmm().

On Mar 6, 2023, 5:45 PM -0300, Viechtbauer, Wolfgang (NP)
<wolfgang.viechtbauer at maastrichtuniversity.nl>, wrote:

Hi Arthur,

Just a small correction: vcov() should be vcalc().

But to your actual question: rma.glmm() doesn't handle that. Some options:

1) use rma.mv() with a measure like "AS" and use vcalc() to construct the V
matrix.

2) go straight to lme4::glmer(). Except for the non-central hypergeometric model,
rma.glmm() is in essence just a wrapper for lme4::glmer() (or GLMMadaptive /
glmmTMB as alternatives).

Best,
Wolfgang

-----Original Message-----
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On
Behalf Of Arthur Albuquerque via R-sig-meta-analysis
Sent: Monday, 06 March, 2023 21:17
To: R meta
Cc: Arthur Albuquerque
Subject: [R-meta] Rare dependent variable with correlation among effect sizes

Hi all,

Tl;dr: I want to meta-analyze studies with a rare dependent variable with
correlation among effect sizes.

I have four randomized controlled trials. Within each RCT, there is one ?control?
group and multiple (>3) ?experimental? groups. Thus, there is a shared control
group which induces correlation among the effect sizes within each RCT.

I am aware that constructing a variance-covariance matrix with vcov() then
fitting the model with rma.mv() is an appropriate solution (per topic 5 in
?Details? in ?vcov). Such approach requires one to first estimate effect sizes
with escalc().

However, I am dealing with RCTs with a rare dependent variable. In these cases,
using an exact likelihood (in this case, Binomial) is preferable. I believe
rma.mv() does not support such likelihood.

How can I fit such model with rma.glmm() considering?correlation among effect
sizes? Ideally, I?d like to fit a random effect model.

Best,

Arthur