thanks for sharing the link, actually the two cases are very similar. Both on survey’s preferences and I guess this is the reason why both of us come out with nested Dirichlet dists. I will have a deep look at his case.
All the different \omega have a meaning that I am trying to model, in terms of priorities of choices from survey responders, and in this sense, they should all be normalised. This is the reason for the hierarchy of Dirichlet dists. But I might be wrong and this is actually not a good way of modelling it.
the costrain is either$\sum \omega =1$ or \sum e^\nu = 1 for all the \omega / \nu (they are actually sort of the same variable)
At the moment, with TruncatedNormal and fixing the error of the dimensionality, it gets to a better fitting, but still, I agree everything seems a bit odd and there is something I am missing
