[R-meta] phylogenetic information in both moderator and random part of rma.mv?
Dear Sigurd, I do not know enough about the specifics of the application to say whether comparing the model with versus without phylogeny is sufficient to conclude something about evolutionary divergence. However, let me make a few comments (I am also basing some comments here on what you wrote to me initially in an email before redirecting you to this mailing list for further discussions): So you fitted a mixed-effects model with lme() of the form: lme(Y ~ Class, random = ~ 1 | species, data=Final.data, method="ML") (or maybe including weights = varFixed(~ sampvar)) and found 'Class' to be relevant (regardless of whether this means: large coefficient, statistically significant, sufficiently larger value of the information criterion compared to a null model). Then you fitted the model below (where you are accounting for phylogeny) and now the support for the relevance of 'Class' disappears or is considerably weaker. I hope this summarizes the issue. First, I would ask: Have you compared the lme() model with this? rma.mv(Y, sampvar, mods = ~ Class, random = ~ 1 | species, data = Final.data, sparse = TRUE, method = "ML") Note that this is not exactly the same model as what lme() fits as explained here: https://www.metafor-project.org/doku.php/tips:rma_vs_lm_lme_lmer This aside, it is quite possible that the relevance / effect / significance of 'Class' changes when accounting for the phylogeny, because this accounts for the potential dependence of the outcome due to similarities between species in a different way than just including species itself as a random effect. Now does it make sense to include Class as a moderator while also including random effects for species? I would say yes. Class is a broader category, so the species random effect accounts for heterogeneity within classes (and the fixed effect for class allows the average of all species belonging to the same class to differ from that of other classes). And whether the values of this random effect are correlated or not (as predicted by the phylogeny) is an empirical question. So if the model with phylogeny fits better, then I would go with that. Best, Wolfgang
-----Original Message-----
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On
Behalf Of Sigurd Einum
Sent: Wednesday, 28 September, 2022 21:04
To: r-sig-meta-analysis at r-project.org
Subject: [R-meta] phylogenetic information in both moderator and random part of
rma.mv?
Dear list,
I want to test for evolutionary divergence (among ectothermic animals) in a
single trait Y. My first approach to this was to test for divergence among
taxonomic classes (specifically amphibians, reptiles, insects, fish, and
crustaceans). I compiled data on several species per class, and multiple
estimates of Y per species (from different experiments), and analysed the data
using a traditional mixed effects model (lme), with species as a random effect
and taxonomic class as fixed effect.
However, one reviewer suggested I should control for phylogeny (without being
more specific). So I built a tree for these species (using package rotl), made it
ultrametric (using compute.brlen in package ape), and computed the variance-
covariance matrix A from this (using vcv from package ape). I then created a
variable phylo to distinguish the phylogenetic component from the non-
phlylogenetic species random effect, and used metafor to fit the model:
rma.mv(Y, sampvar,mods = ~ Class, random = list(~ 1 | species, ~ 1 | phylo),
R = list(phylo = A), data = Final.data, sparse = TRUE, method =
"ML")
(sampvar is the sampling variance associated with each estimate of Y)
However, now I have started doubting whether this model makes sense, i.e. to
estimate an effect of taxonomic class (which in essence is a phylogenetic effect)
while simultaneously modelling a random effect of phylogeny. Would an appropriate
alternative be to not have class as a moderator, but rather compare fits of
models with and without the phylogenetic variance-covariance matrix. If the model
including phylogeny is better than one without it, can I conclude that there is
evolutionary divergence in this trait?
Any advice anyone might have on this would be much appreciated!
Best wishes
Sigurd Einum