Default for thresholds in brms for cumulative link models

Dear @paul.buerkner and dear members of the community

we are working on a model with ordinal response data and we are using your brms package.

We would have two questions concerning how the default priors for thresholds are implemented.

  1. We are aware that student_t(df, m, sd) distributions are used, but how are the parameter values selected?

  2. How is the ordering of the thresholds guaranteed?

Thank you very much

I believe that (@scholz might have something to say):

  1. brms peaks at the data I believe. In short, never use default priors. Do prior predictive checks.
  2. Use factors?

I haven’t worked with ordinal models so sadly can’t help with this.

  1. The student-t distribution is chosen in an rather adhoc manner. 3 DF ensures that the tails are very wide while still having mean and variance. The location (m) and scale (SD) is chosen based on the link transformed mean and SD of the data. The latter is not applicable in ordinal models, so set to zero and 2.5 by default which should be weakly informative for the link functions used in ordinal models in brms.
  2. via the ordered class in Stan.