Hi @cluhmann thanks for reaching out!
It took me a while to come up with something more complicated, and in the end, I just decided it’s best to work with a setting as close as possible to my data. I simulated some data and applied my routine, which you can find in a gist here: https://nbviewer.org/gist/nmwitzig/3fa7617df49220f4e22cb02f2dea4268
We see that even with simulated data, the chains don’t converge and do not fully recover the parameters of the first stage (which I called the “performance measurement phase”); What makes things worse is probably that I observe only a single observation per person in the second stage (which I try to predict based on the performance measures).
Let me know if you need more! I know it’s a bit complicated setup in the Gist, but the modeling part in itself is (hopefully) quite straightforward.
Thank you for your time and patience!
Best
N