Thanks, I understand now.
I’ve been playing with missing values myself, and I recently discovered this paper:
Bayesian PCA gets good results. It would be great if someone implemented this for the PyMC3 documentation.
You might be also interested in this library: