Issues with parameter recovery

If you think it’s the hierarchy that is driving these pathologies, I would encourage your to investigate the sampling itself and not necessarily the accuracy of your parameter recovery. For example, the decay parameter may not be all that identifiable and may exhibit strong dependencies with other model parameters. This notebook (which you may be familiar with) provides some examples of how to do this in the case of hierarchical models. On top of that, I would not assume that a non-centered parameterization will necessarily be better. It is not a universal truth that non-centered is superior to centered parameterizations.