Can you use a horseshoe prior with a categorical model?


I’m quite new to brms so apologies if this is a naive question. I’m trying to fit a multinomial model each datapoint belongs to one of three categories I’m trying to predict which with c. 10 predictors.

I suspect most of these have little effect (which I can see when I fit the model without trying to remove predictors). I could just report this maximal model but it feels like it takes up unnecessary space and would like a clear statistical support in identifying which factors are most influential.

However, when I try to use a horseshoe prior like so.

Response ~ X1 + … + X10 + (1 | id) ,
iter = 4000,
warmup = 2000,
prior = set_prior(horseshoe(1)),
thin = 2,
data = ABC,
family = categorical(link = “logit”),

I get this message

“Specifying global priors for regression coefficients in categorical models is deprecated and may not work as expected.”

It doesn’t trigger any warnings with fit and the results make sense with the previous model with just weakly informative gaussian priors on the predictors. Is it safe to ignore this warning? I understand the logic behind it in trying to ensure all priors are properly considered, but in this case is it necessary?

  • Operating System: Windows 10
  • brms Version: 2.13.0

Tagging @jpiironen and @avehtari.

The warning just tells you that this way of specifying priors still works but may not work in the future at some point .