I wanted to do a multilevel meta analysis and I typed this command

model5<-rma.mv(yi, vi, mods=~factor(intcode)+basemf-1,     random=~1|stdy/id, data=salt) 

where study is level 2 and id are communities (level 1) nested within level2. I have 18 studies with 28 ids (communities). I got the following output

Multivariate Meta-Analysis Model (k = 28; method: REML)
Variance Components:

    estim    sqrt  nlvls  fixed   factor
sigma^2.1 0.0004 0.0200 18 no stdy sigma^2.2 0.0016 0.0406 28 no stdy/id
Test for Residual Heterogeneity: QE(df = 24) = 912.9175, p-val < .0001

Test of Moderators (coefficient(s) 1,2,3,4): QM(df = 4) = 124.2892, p-val < .0001

Model Results:

                estimate   se     zval    pval    ci.lb    ci.ub     
factor(intcode)1 -0.0360 0.0144 -2.5028 0.0123 -0.0643 -0.0078 *         
factor(intcode)2  0.0482 0.0263  1.8323 0.0669 -0.0034  0.0998 . 
factor(intcode)3 0.0019 0.0348 0.0551 0.9560 -0.0662 0.0700
basemf 0.0046 0.0007 6.8576 <.0001 0.0033 0.0060 ***

Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.

> confint(model5) [[1]]

        estimate  ci.lb  ci.ub
sigma^2.1 0.0004 0.0000 0.0028 
sigma.1 0.0200 0.0000 0.0534


          estimate  ci.lb  ci.ub
sigma^2.2 0.0016   0.0006 0.0039 
sigma.2   0.0406   0.0250 0.0621

The output indicates that intervention 1 and basemf are significant predictors of proportion infected after 1 year of intervention. The variance components indicate the variablity between and within studies are very low. I have the follwing queries How do i interpret these findings? Is it possible to estimate a pooled effect size for this model? How do I get the values of I^2 as i need that to compare across models (random effects, mixed effects, multilevel random effects Vs multilevel mixed effects). As there are only 18 studies, i cannot add anymore predictors....so What should i do next? Regards Vidya



You can obtain $I^2$ like statistics by following the advice given in the post computing heterogeneity assigned to random factors in meta-analysis

It looks to me as though your variances are rather poorly estimated and I would strongly recommend using the profile command on your model to check the shape of the (pseudo)-likelihood. If it is flat it suggests that you do not have enough data to identify the model well. I may be wrong here of course and everything may be OK.

I agree with your statement that with that number of studies your ability to search for moderators is going to have to be limited. You can obtain a test of your factor moderator as a whole rather than relying on the test for each individual coefficient you need the btt parameter


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.