Thanks, @junpenglao, good shout.
I am trying to build a regression model for \nu from X∙ν_c=y where c is the index for the hierarchical model.
X has a 1 and a -1 on each row places sort of randomly, the rest is zeros
X=\left\lceil \begin{matrix} 0 & 1 & -1 & 0 & 0 & 0 \\ 1 & -1 & 0 & 0& 0 & 0 \\ 0 & 0 & -1& 0 & 0 & 1 \\ & & ... & & & \end{matrix} \right\rceil
the y values can only assume 4 values from y=log(2d+1) with d∈\{1,2,3,4\}
and d=1 is more frequent than d=2 and so on, and these two things give the “weird” curve on the observed line in the previous picture, with the 4 picks corresponding to the 4 values of d
Finally, the Dirichlet distribution of \nu is it for a normalisation factor such as \sum_{}^{}e^\nu=1 , but I don’t think it gives problems
The model that gave me the result of the posterior predictive is the one below, but it gives a high error for the long tail on the negative side. I have tried other distributions but nothing, getting NaN errors and now I am trying now with the Multinomial to have a better fitting…
Also, I don’t know why the output has (85x85) dimension (85 is the length of y ) should it not be just (85,)?