Hi Aki,
I thought that since the family and group-level term were constant across models, they wouldn’t affect model weights. Now that we know I have misspecification, they’re clearly important. Thank you for your patience - I’m still very new to Bayesian stats.
The phylo term accounts for the fact that the species/samples are not phylogenetically independent. I followed this vignette to set it up. I am not sure what kind of prior I’d assign to it.
I did run the full model with the beta family and better thinning/warm up numbers. Here’s how I specified it for completeness:
LOO R2 = 0.0521, so it’s still very bad at explaining the data.
The posterior predict plot looks like this:
Is this sufficiently well specified to state that the model isn’t really predictive, and that phylogeny is the only thing with any signal at all? Either way, what do you look for to feel confident about your models’ specifications?