Thanks again to @mattansb for the detailed replies. Thanks also to the other responders in this thread.
I’ve now worked through a variety of related examples (not posted), and have come to this conclusion:
When the predicted variable is categorical (named resp in a matrix of counts) and the predictor is categorical (named group), the most robust[1] approach in brms is
formula = resp | trials(n) ~ 0 + group for non-hierarchical (suppressing the intercept with ~ 0 + makes it easier to specify the prior — this was not featured in any of the examples),
formula = resp | trials(n) ~ 1 + (1 | group) for hierarchical,
both with
family = multinomial(link = logit, refcat = NA) (yes, that’s NA for refcat) and a (moderately) informed prior. Check that the prior is really doing what you intend.
The approach I proposed, using a dummy reference category, also seems to work pretty well! But it introduces subtle symmetric effects that have not yet been thoroughly characterized (@mattansb referred to it as a form of regularization).
Thanks again to all!
By “robust” I mean not introducing asymmetries across response categories, mathematically clear, and reasonable execution time. ↩︎