The hierarchical model in the notebook has clearly not converged, so if your samples look like that I wouldn’t trust much any of its results. A Primer on Bayesian Methods for Multilevel Modeling — PyMC3 3.10.0 documentation should have the centered version of that same model and it converges.
There are also other things to update and improve, I’ll keep track of those at GLM hierarchical · Issue #80 · pymc-devs/pymc-examples · GitHub in case anyone seeing this is interested in updating the notebook.