# Metafor package: choice of moderator for bias correction

I am doing a 3-level meta-analysis using rma.mv() function. I found a thread about bias diagnostics with an advice to use "regression test for funnel plot asymmetry". I have two questions:

1. My outcome variable is log odds ratio. Which moderator should I use: sample size, or sampling variances, or inverse of sampling variances?
2. Нow to interpret the results of these regressions? Do they show the presence/absence of the bias, or do they correct bias?

I tried two different moderators and got totally different results:

• model with sample size as moderator -- slight change in mean effect size; test of moderators non signif.

rma.mv(yi, V, random = list(~1 | ID_2, ~1 | ID_1), mods = ~ sample_size, data=data)

• model with sampling variance - huge change in mean effect size; test of moderators highly signif.

rma.mv(yi, V, random = list(~1 | ID_2, ~1 | ID_1), mods = ~ V^2, data=data)

• What happens if you use the inverse of the sampling variance or one of the other options outlined in the answer you cite. Note that this issue has also been discussed on the R mailing list for meta-analysis stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis/ so you might find the archives there helpful. – mdewey Jun 28 '20 at 12:37