# Running a Multilevel Linear (Mixed-Effects) Meta-Analysis Model without a variance-covariance matrix

I have a dataset that I want to run a meta-analysis on. However, the data has a hierarchical structure with several endpoints per study and where studies are further clustered together. So, I should run a multilevel model to account for this.

Reading the metafor manual pages for rma.mv reveals that you can input a

vector of length k with the corresponding sampling variances or a k×k variance-covariance matrix of the sampling errors.

In my case, I don't have the variance-covariance matrix and there's absolutely no way for me to find out what it is, so I'm just supplying the function with a vector of sampling variances.

Now, I'm trying to understand what practical effect this will have on the outcome of my analysis. Can/should I still run the model or do I need a variance-covariance matrix in order for it to make sense?

• Try mvmeta. You can specify a given correlation (eg 0.5), and then run sensitivity analyses with other specifications. – Joe_74 Feb 15 '19 at 10:05
• Try asking on the R mailing list stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis/ but browse the archives first as this type of issue has come up repeatedly. – mdewey Feb 15 '19 at 13:02