Thanks a lot for this! n_eff and Rhat look beautiful and it took about half the time to run. Still getting quite different results with variational inference though. However, I think I should do some more careful checks of my test and results before taking this too seriously.
Indeed this is supposed to be a latent factor model, just for categorical variables. I just found a couple of useful looking discussions on identification - a blog post and a related discussion on this forum. So I’ll have a read and try and make some improvements.
Many many thanks!