I want to build a hierarchical model with let’s say 5 submodels and let’s say 5 stochastic RVs per submodel. I want to try a multivariate normal prior on all of 25 lower-level RVs, but I’m worried about funneling, so I want to use a non-centered parameterization. Has anyone written elegant PyMC3 code for this (a noncentered multivariate normal hierarchical model) already? It would be great not to reinvent the wheel!
Most of the models in the updated radon NB are non-centered, and there is a new MvNormal one
So I think you’ll find it useful
Did you write/update that notebook? I ask because I can’t see intuitively how that non-centered parameterization solves the funneling problem (in my use-case, I’m still getting divergent samples, centered or non-centered).
Yep, I wrote the updates of the different models – the ArviZ magic was done by the great @OriolAbril.
This can be due to a wide number of things, unfortunately. I’d check that priors are in line with your domain knowledge and regularizing enough (very important in hierarchial models), and if there is any strong correlations between parameters’ posteriors.
Also, to make things even more interesting (), if the posterior of the population sigma (the one which induces shrinkage of parameters) is far enough from zero, the centered parametrization will usually be better suited.
Hope this helps
Just wanted to say thanks for the help! My non-centered multivariate hierarchical model seems to be working well. I did have to play with
target_accept quite a bit, though.
Thanks for the feedback, really glad to know when a resource is useful!