Thank you for the informative responsive–you’ve given me a few things to think about. I tried compile_kwargs={‘mode’:‘FAST_RUN’} but didn’t get any improvements. I will have to do some reading up on ADVI and pathfinder. I do have a few questions if you wouldn’t mind:
- When you say custom HurdleGamma, do you mean implementing a sort of two-step model that has a Bernoulli or Binomial and then a Gamma? I had thought of this somewhere along the way, but the HurdleGamma seemed like the easier option for implementation (until I hit this wall). Seems to me at this point like this option would be the path of least resistance.
- I know very little about Bambi right now, but does it use the same HurdleGamma as pymc? If so, how are you getting past those issues with sampling? Or, is it just that you are using a more manageable amount of data?
- Is there a sampler of those 4 you listed which you’ve found to be most suitable for your use cases?
I’ve attached a sample dataset here with the following:
record: unique to every row
group: each group ID has ~30-50 records
observed_value: observed data for each record
I really appreciate your help!
hurdlegamma_sample.csv (35.5 KB)