[R-meta] Quick question about the metafor package: log-odds with rma()

[Viechtbauer Wolfgang (SP)]

Assuming the studies are not small, it will probably make very little difference whether you pool the (log) odds ratios via a meta-analysis or whether you pool the values by using a combined dataset (using study as a fixed/random effect). The advantage of using the raw data (if available) is that one can examine subject-level predictors (e.g., the age of each subject) while a 'standard' meta-analysis would only allow you to examine study-level predictors (e.g., mean age). But if one isn't interested in subject-level predictors, then power is more or less the same (for certain outcomes, like the mean difference, it is the same).

> > Given this, in R, I am using the metafor function rma(yi, vi, ...) where
> > yi=estimates of log-odds, and vi=standard errors ensuing from the glm
> > models, using the Hartung-Knapp correction. And I would like to know if
> > using log-odds as effect size estimates is appropriate.

The second argument of rma() (called 'vi') is for the *sampling variances*, not the standard errors. If you have the standard errors (square root of the sampling variances), you can use rma(yi, sei=SE), assuming the variable with the standard errors is called 'SE'.

That aside, yes, you can just pass log odds ratios and corresponding sampling variances (or standard errors) to rma(). For once, it is that simple.

Best,
Wolfgang

-----Original Message-----
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On Behalf Of Flávio Azevedo 
Sent: Sunday, 17 September, 2017 19:18
To: James Pustejovsky
Cc: r-sig-meta-analysis at r-project.org
Subject: Re: [R-meta] Quick question about the metafor package: log-odds with rma()

Dear r-sig-meta-analysis,
(and James and Wolfang)

Thank you for your kind and explanatory answer James. And apologies for
addressing it to Wolfgang, I now understand how​ ​this​ ​works.​

​I will proceed with your suggestions and also estimate the log-odds we are
interested based on the pooled data.

Lastly, I leave the below information in case it changes your prior
suggestion. The data ensue from 4 surveys/studies in Social Sciences,
mostly political variables. Three samples used large convenience samples,
and 1 is representative w.r.t. the target population via a professional
survey company. All estimates ensue from Logistic regressions with the
form: glm(DV ~ IV + same.controls, family​ =​ ​binomial(link='logit').

All the best,

\Flavio Azevedo

On 17 September 2017 at 17:58, James Pustejovsky <jepusto at gmail.com> wrote:

> Flavio,
>
> (Recognizing that your question was addressed to Wolfgang, I will offer a
> response anyways.)
>
> I think it is hard to judge whether the analysis you propose is
> appropriate or not, without having more information about the variables
> involved and whether they are measured in consistent fashion across the 4
> studies in your analysis.
>
> The analysis you describe is not inappropriate--and indeed it is probably
> a good thing to run as a first step. However, if the four studies are
> indeed close replications, you might want to consider pooling the raw data
> from across the four studies and analyzing the combined data as you would a
> block-randomized trial. I think this would likely provide more accurate
> estimates and more powerful tests of the focal effects, although again it
> is hard to say for certain without knowing more details.
>
> James
>
> > On Sep 17, 2017, at 9:26 AM, Flávio Azevedo <falafla at gmail.com> wrote:
> >
> > Dear Dr. Wolfgang Viechtbauer,
> >
> > I am a Ph.D. candidate at Cologne University and I am conducting a
> > mini-meta analysis (within paper effects) of log-odds ensuing from 4
> > studies of a given DV on a set of continuous and categorical covariates.
> > All models are equal across the 4 studies.
> >
> > Given this, in R, I am using the metafor function rma(yi, vi, ...) where
> > yi=estimates of log-odds, and vi=standard errors ensuing from the glm
> > models, using the Hartung-Knapp correction. And I would like to know if
> > using log-odds as effect size estimates is appropriate.
> >
> > My question stems from the simplicity of this approach, which in
> statistics
> > is almost never the case. Thank you very much for your time and work on
> > this,
> >
> > All the best,
> >
> > \Flavio Azevedo