On thing that hasn’t come up is looking at posterior pair plots. This is my first step when diagnosing divergences.
That’s an excellent point and in fact the original reason for this thread. Indeed the pair plots were my first go-to but since I have 3000+ groups I couldn’t figure out how to actually inspect them. Clearly doing so for 3000+ pair plots wouldn’t have been feasible and I hoped that there is away of somehow “sorting” the groups by how problematic they are or otherwise drawing my attention to a hopefully smaller subset of the 3000+
In the end that was unnecessary when I realised that the clipping was at fault, but the original problem still stands. Especially since Betancourt advises that sometimes you might want to have some of your groups parametrised centrally and others non-centrally. If one were in that situation with 3000+ groups it’d be a challenge