Hi,
- Do you get some warning / divergences from NUTS after sampling? This is usually very helpful to understand what’s going wrong.
- Initializing your sampler at the MAP is pretty much discouraged now.
- Have you done some prior predictive checks? The prior stds on your hierarchical means,
mu_a
andmu_b
seem super huge. With hierarchical models, implementing regularizing priors is usually very important.
Hope this helps