I’m trying to model some ordinal response data (Likert data) with repeated measures. I’d like to use a random effects model with brms to capture variation among questions and respondents, etc. However, I find I cannot completely reduce my divergent iterations to 0.
I want to follow the advice of Michael Betancourt in this tutorial. Specifically, he recommends using a Dirichlet prior for the cutpoints. As I understand it, this is handy because it can be very hard to estimate the cutpoints when there are very few observations in one category (e.g. almost nobody scores something as 1 out of 7).
It seems like
brms uses normal priors for each cutpoint, and allows random effects. Meanwhile
rstanarm::stan_polr uses a Dirichlet prior for cutpoints, but no random effects. Is it possible to have both? Or does this require adapting Michael’s code and writing Stan directly?