Interpretation of cor term from multivariate animal models

[ttburk]

@martinmodrak

That’s basically what I was interested in. But still kind of hard to guess the cause. Just to fiddle around a bit more - what if you use both responses, but play with the correlation term (I would try at least (1 || ID), (1 |p| ID) and (1 | gr(ID, cov = A))) which of those give you divergences?

…They all give me divergences haha.

@martinmodrak

The problem IMHO is (as I think @diogro alluded to) that you need (1 | p | ID) term to catch phenotypic correlations (assuming those are random among the population, not correlated with genes). If you only include (1 | p | gr(ID, cov = A)), the correlation coefficient might “soak up” some of the phenotypic correlation, as there is no other part of the model where it could be reflected.
[…]
The problem is that having both (1 | p | gr(ID, cov = A)) (genetic correlation) and (1 | p | ID) (phenotypic correlation) in the same model sounds like a recipe for fitting issues unless you have tons of data (and maybe even if you have tons of data). Moreover, you already have fitting issues when using just one of those terms and adding more terms rarely makes a model better behaved…

EDIT: When using two correlation terms it would be necessary to let them differ, so the hypothetical full formula would be Sex + Site + (1 | p1 | gr(ID, cov = A)) + (1 | p2 | ID).

Ok–I think I understand what you and @diogro are talking about now with the phenotypic and genetic correlation terms, thanks for clarifying! However, it looks like I can’t model a repeated group-level effect?
brms version 2.13.0
Error: Duplicated group-level effects are not allowed. Occured for effect 'Intercept' of group 'ID'.
Looks similar to a previously posed (unsolved) question here.

As the general consensus seems to be, the data seem too few to be able to fit these models really well. An alternative I will consider if I cannot build an appropriate bivariate model would be to report how large a sample I would need for this analysis to be possible, and I’ll use advice from this post as a starting point to understanding how to do this in a Bayesian framework.