Ordinal probit response model (also translation from jags)

[rongmu]

I’m on the same boat. The implementation of the thresholds is causing sampling issues.

nhuurre offers a solution that works very well (check the link below). The solution is even better than the original DBDA2e parameterization IMHO, as it ensures that the thresholds are ordered.

Sampling issues: Ordinal probit model with fixed threshold extremes

Here’s how to properly construct an ordered vector with fixed upper and lower bounds parameters { ... simplex[K-2] c_; } transformed parameters { ordered[K-1] c; // thresholds c[1] = c_lower; for (i in 2:(K-1)) { c[i] = c[i-1] + c_[i-1]*(c_upper-c_lower); } ... } Second, that uniform prior on sigma is going to cause problems. I’d say try normal(0, K) instead (it is automatically truncated to positive only). And finally, don’t compute theta manually. Stan has a builtin order…