Thanks. I appreciate the feedback.
The “parallel/hand-coded” model should ultimately be the same as the cleaner pymc-indexed method. Unless I’m misunderstanding something. A key motivation for posting the question.
I coded things up by hand in an effort to diagnose the oddities I’m observing in the latter cleaner approach. The hand-coded approach is essentially the same as [this](https://mc-stan.org/docs/2_21/stan-users-guide/hierarchical-logistic-regression.html) example in the Stan documentation.
The idea is that we groupby on the data by group to get the group name and group data, and then create the parameter distributions for each group independently, calculate the logits for the group. Then in the end, we just concatenate, invert, and infer.
Good point regarding the target_accept and max_treedepth. These were set during an earlier iteration, and don’t seem to be required now, but even if they are, how can I identify the issues with the latter specification, and why is it different than the former “hand-coded” specification?