By “VI” you mean “variational inference,” correct?
I’m not sure what exactly is going wrong here, because I’m testing the logp of samples generated using sample_prior_predictive, so there isn’t a VI stage here, the way there is when I use sample with NUTS, is there?
I don’t know how to check the internals of the model to explain this, except to explain that it is a three-level hierarchical model, because it involves measurements of populations, and that the population measures (per other question) is the sum of independent Gaussians. I’ll see about putting the model into a notebook to share.
Is there any chance that the logp simply decays as the number of variables in the model grows, and that beyond a certain number of variables we need to rescale?
Is there any tracing mechanism that would let me take the logp calculation for a single point and print out all of the factors? That would show us if it’s decaying with size as I suggest.
Thanks for any suggestions!