[R-meta] Subgroup analysis with RVE
Viechtbauer, Wolfgang (SP)
wo||g@ng@v|echtb@uer @end|ng |rom m@@@tr|chtun|ver@|ty@n|
Thu Dec 3 12:30:30 CET 2020
Interesting, thanks for this clarification on terminology.
In principle, it is an arbitrary distinction anyway, since one can easily fit a meta-regression model that allows tau^2 to differ across subgroups, which is exactly identical to fitting RE models in the two subgroups.
But this is only true when the only predictor is the factor distinguishing the subgroups. Once the model includes additional predictors, the equivalence breaks down (unless one fits a model that allows the coefficients for the additional predictors to also differ across subgroups).
>From: Dr. Gerta Rücker [mailto:ruecker using imbi.uni-freiburg.de]
>Sent: Thursday, 03 December, 2020 12:15
>To: Viechtbauer, Wolfgang (SP); Ioana Cristea
>Cc: r-sig-meta-analysis using r-project.org
>Subject: Re: [R-meta] Subgroup analysis with RVE
>Dear Wolfgang, dear Ioana,
>This is only to hint to a different use of the notion "subgroup
>analysis" at Cochrane, which may have led to misunderstandings here.
>Cochrane uses "subgroup analysis" to denote metaregression with a
>nominal covariate (common heterogeneity parameter and between-groups
>test), not for separate analyses within groups. See the Cochrane
>Handbook for Systematic Reviews of Interventions,
>. Thus, Cochrane's subgroup analysis is a special case of
>metaregression. This notion is also used in R package meta.
>Am 03.12.2020 um 10:10 schrieb Viechtbauer, Wolfgang (SP):
>> Yes, adding this variable as a covariate is indeed meta-regression with a
>dichotomous predictor. This assumes that the amount of heterogeneity is the
>same within the two subgroups. If you subgroup, then you automatically allow
>the amount of heterogeneity to differ between the two groups. This is
>> When combining this with RVE, things are a bit different though. RVE
>doesn't make assumptions about the underlying variance structure (in fact,
>the whole point of RVE is that it works (asymptotically) even if the var-cov
>structure is misspecified). So, even if tau^2 differs across the two groups,
>RVE in the context of a meta-regression model is still going to provide
>>> -----Original Message-----
>>> From: Ioana Cristea [mailto:ioana.alina.cristea using gmail.com]
>>> Sent: Thursday, 03 December, 2020 9:31
>>> To: Viechtbauer, Wolfgang (SP)
>>> Cc: r-sig-meta-analysis using r-project.org
>>> Subject: Re: [R-meta] Subgroup analysis with RVE
>>> Thank you, sorry I did not explain clearly. I estimated effects within
>>> subgroups with RVE (intercept only model in each), but I did not do a
>>> between-groups test of significance. I did fit a model with a between
>>> covariate (subgroups coded dichotomously: high/low), which I took to be
>>> equivalent of a meta-regression with a dichotomous predictor.
>>> On Thu, Dec 3, 2020 at 9:20 AM Viechtbauer, Wolfgang (SP)
>>> <wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:
>>> Dear Ioana,
>>> I am not sure why you think that a test of that moderators is not useful
>>> when the number of effect sizes is so different between the two groups.
>>> else equal, power will indeed be lower as opposed to the case where k is
>>> similar for the two groups, but that will be the case no matter how you
>>> analyze the data (via meta-regression or subgrouping).
>>> This aside, I am a bit confused by your question. A subgroup analysis is
>>> just that: Fitting a particular model in a subgroup of the studies. That
>>> be done with or without RVE. You seem to have done this already ("I
>>> estimated effects in the high and low subgroup separately").
>>>> -----Original Message-----
>>>> From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-
>>>> On Behalf Of Ioana Cristea
>>>> Sent: Wednesday, 25 November, 2020 15:43
>>>> To: r-sig-meta-analysis using r-project.org
>>>> Subject: [R-meta] Subgroup analysis with RVE
>>>> Dear all
>>>> I am running a meta-analysis with robust variance estimation (RVE) and
>>>> focus is on a categorical variable (a risk of bias domain variable,
>>>> dichotomously, high vs low, in the database "0" for low risk and "1" for
>>>> high risk). I estimated effects in the high and low subgroup separately,
>>>> each in an intercept only model. I also ran the RVE model estimating the
>>>> between-study effect of covariates, including my predictor (which is a
>>>> between study variable) as a covariate (it was not significant). I think
>>>> big problem is that the two subgroups delineated by the categorical
>>>> variable are very uneven in terms of the information contained (one has
>>>> effect sizes, one 13) and in my view a test of significance for
>>>> between them is not useful.
>>>> One of my co-authors would also like us to run a subgroup analysis. I am
>>>> not sure how to do that, if it is possible with RVE and if it adds
>>>> to the model estimating between-study covariates, which I already ran.
>>>> Thank you!
More information about the R-sig-meta-analysis