I have a multivariate hierarchical model with very few observations (2 per cluster). Consequently, the model takes hours to be fitted, with a lot of divergent transitions and low n.eff. ratios in the coefficients.
But my biggest problem is that it produces predictions that are totally out of scale and impossible (my focus is prediction).
I’m trying to fiddle with priors to fix this, but, is there a way to put priors on the outcome possible values and let all the rest of the sampling be dictated by this? that is, discard features’ values that allow such extreme values?