Viechtbauer, Wolfgang (SP) wo||g@ng@v|echtb@uer @end|ng |rom m@@@tr|chtun|ver@|ty@n|
Wed May 5 16:38:54 CEST 2021

```If you mean something like this:

dat <- dat.konstantopoulos2011
res0 <- rma.mv(yi, vi, random = ~ 1 | study, data=dat)
res1 <- rma.mv(yi, vi, random = ~ 1 | district/study, data=dat)
anova(res1, res0)

then yes, since in this case the null hypothesis says that H0: sigma^2_district = 0, which is at the boundary of the parameter space. Whether a 50:50 mixture of a chi^2 random variable with df=1 and a point-mass at 0 is still the correct null distribution is a separate issue.

However, I would personally not bother testing this or at least not modify my model based on the results of such a test. If the data have a multilevel structure, then I would analyze them accordingly. If the estimate of sigma^2_district should turn out to be small, then so be it.

Best,
Wolfgang

>-----Original Message-----
>Sent: Wednesday, 05 May, 2021 16:27
>To: Viechtbauer, Wolfgang (SP); r-sig-meta-analysis using r-project.org
>
>Thank you Wolgang for the clarification
>So If we test for the need for 3-level or 2-level MA, in that case we'll use a
>mixture of chi-square?
>
>On 04/05/2021 12:59, Viechtbauer, Wolfgang (SP) wrote:
>You can do a likelihood ratio test. Using the same example:
>
>library(metafor)
>dat <- escalc(measure="RR", ai=tpos, bi=tneg, ci=cpos, di=cneg, data=dat.bcg)
>dat\$alloc <- ifelse(dat\$alloc == "random", "random", "other")
>res1 <- rma.mv(yi, vi, mods = ~ alloc, random = ~ alloc | trial, struct="DIAG",
>data=dat, digits=3)
>res0 <- rma.mv(yi, vi, mods = ~ alloc, random = ~ alloc | trial, struct="ID",
>data=dat, digits=3)
>anova(res1, res0)
>
>The issue you mention at the end is not relevant here, since we are testing H0:
>tau^2_1 = tau^2_2, not something like H0: tau^2 = 0. In the latter case, the
>parameter is at the boundary of the parameter space under the null hypothesis,
>which leads to the issue you mention.
>
>Best,
>Wolfgang
>
>-----Original Message-----
>Sent: Tuesday, 04 May, 2021 12:52
>To: Viechtbauer, Wolfgang (SP); r-sig-meta-analysis using r-project.org
>
>Thanks Wolfgang!, simple and elegant explanation in your post.
>
>Is it possible to check which assumption fit better the data (same variance vs
>different variance in each subgroup)?
>
>The option "different variance" is more general than "same variance". So, is it
>possible to say that they are nested and do an ANOVA between them?
>
>I wonder if for such "variance test" the usual chi2 distribution doesn't apply,
>and require a mixture of chi2, as it has been propose previously when the test for
>RE variance are conducted.
>
>On 04/05/2021 12:14, Viechtbauer, Wolfgang (SP) wrote:
>If one runs separate meta-analyses, one can also test for subgroup differences.
>This is not a distinguishing characteristic. The main difference is that separate
>meta-analyses automatically allow all parameters (including any variance
>components) to differ across analyses, while a single meta-regression model (with
>a categorical moderator) will by default assume that all parameters except of
>course for the subgroup means are the same across subgroups. But even this
>assumption can be relaxed and one can fit a meta-regression model that will give
>you exactly identical results as fitting separate meta-analyses within subgroups.
>See here:
>
>https://www.metafor-project.org/doku.php/tips:comp_two_independent_estimates
>
>The same idea generalizes to models such as those that can be fitted with
>rma.mv().
>
>Best,
>Wolfgang
>
>-----Original Message-----
>From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces using r-project.org] On
>Sent: Tuesday, 04 May, 2021 12:04
>To: r-sig-meta-analysis using r-project.org
>
>Thanks Gerta for such a simple and important reminder.
>
>Apart from having test for subgroup differences, which other advantage
>can have doing a subgroup analysis (with the moderator in
>meta-regression) vs separate meta-analyses?
>Just assuming that is a categorical moderator
>--
>Kind regards