hi, thanks for the suggestions. They have both helped a lot with speed and stability. The blog-post linked was very informative. I did have two questions regarding it, however, that I haven’t been able to find an answer to.
First, what is the advantage to combining R and beta into beta_tilde?
I encountered and issue with this as often the R matrix on my data is singular, and thus not invertible. Is this a necessary transformation in the model or can you simply model it as: pm.math.dot(Q,pm.math.dot(R,beta)) ?
Second, are any special transformations necessary to interpret the posteriors of the coefficients as log-odds as with normal logistic regression?