I have run a heirarchical model using stan_lmer but despite increasing adapt_delta up to 0.999 I continue to get the same one divergent draw… As suggested, I ran pairs, etc. and indeed I found a funnel. So I read about how to reparameterize stan code (including the famous 8 schools problem). But I would rather not have to figure out how to submit my problem in native STAN code. (After all, that is why I have been using rstanarm.) I also discovered how to recover the stan code from my model, and as you folks have been saying, that model specification is not intended for human reading. One suggestion was to use brms, but that code was only marginally more comprehensible to me.
So I have two questions:
a) Is it possible to obtain the fully compiled stan code (with expansion of all the headers, etc) to look at what is actually sent by rstanarm to STAN? I might be able to work with that.
b) Given that the funnel is a common enough issue with hierarchical models, has anyone considered giving the user the option of submitting a non-centered version of the model? I recognize that for a big enough sample, using the centered code is more efficient. But where one hits a funnel and reparameterization is the only available fix, it would be nice to be able to run a model without divergences.
I would submit the model and the data, but that is confidential.
rstanarm Version: 2.21.1 on R 4.1.2, RStudio 2021.09.0 Build 351 Windows 10 up-t-date.
Thanks for any suggestions.