What does the hierarchical model look like when having missing in observed?

In multivariate case this won’t work, currently the missing value feature in pymc3 only handles 1d case. You can have a look at some related discussion here: OOS predictions with missing input values