[R-sig-ME] Post model fitting checks in Metafor (rma.mv)

[Shona Smith]

Hi all,

I am currently conducting a meta-analysis using rma.mv in metafor.  My model uses Hedges’ d (converted to g) and includes 3 moderators (age-3 levels; treatment-5 levels; biomarker-3 levels).  I have included 3 random effects: species nested within taxonomic class (since I have more than one study for some species, and species are spread over 7 taxonomic classes) and study separately.  So my code is as follows:

rma.mv(yi, vi, mods = ~ Age + Treatment + Biomarker, random = list(~ 1 | Study, ~ Species | Taxonomic.class), data=mydata)

I was wondering what the best method for post model fitting checks was in rma.mv?  I know in the reference manual it mentions profile.rma to create a plot of the restricted log likelihood and I have done so.  However, I wondered if I need to plot all 3 variables (sigma2, tau2 and rho) and also how I know which value to specify for each?  Am I correct in that I should see a clear peak in each graph?  Is there anything else I should be looking for?

For post model fitting checks should I also look at residual normality and residual against fitted values, as would be done for a typical mixed model?  I think the standardised residuals are best for this – I can get them with rstandard.rma.mv, but it does not allow me to plot them. 

Finally, when I include the intercept in the model, I can see if there are significant differences among moderator levels.  However, I was wondering what the output includes when the intercept is not included: is this the overall effect size estimates for each moderator level?

Kind regards,
Shona


Shona Smith
PhD Student

Room 321
Institute of Biodiversity, Animal Health and Comparative Medicine
Graham Kerr Building
University of Glasgow
Glasgow G12 8QQ