Non-centered parameterization of ordered vector

I am estimating a cumulative odds model and I have a set of cut-points/intercepts that are ordered. I am using a t(3, 0, 5) prior, which seems to work quite well. Is it advisable (and more importantly, is it even possible) to use non-centered parameterization in this case?

I have L levels in the response, and there are K groups, each of which has its own cut-points. The current parameterization is

parameters {
ordered[L-1] c[K];      //  (K X [L-1] matrix)

model {
  for (k in 1:K)
      c[k] ~ student_t(3, 0, 5);

I attempted to use the gamma(3/2, 3/2) and standard normal parameterization, but the structured/ordered nature of this parameter doesn’t seem to translate well under this transformation. Again, I am wondering if it is even possible or advisable to attempt this.


Sorry for taking quite long to respond.

That’s not a question that can be easily answered in isolation. Depends on why would you want to use it and what is the rest of the model. Are you seeing patterns similar to the ones at Visual MCMC diagnostics using the bayesplot package • bayesplot ? If yes, non-centering might be sensible.

Also non-centering is used when the scale parameter of the distribution is a parameter. If the scale is fixed then there is usually no benefit from non-centering.

Best of luck with your model!

1 Like