Thanks for the reply.
Can you explain what why for the flat prior? I’m trying to use the asymmetric sampling trick since hierarchical models are so hard to sample. When I run a simulation without cross cutting effects the flats don’t present a problem because of the term where I multiply raw*sigma. I recover my parameter values and the NUTS is less likely to get stuck and never complete when I run it this way.
Now I feel really silly. Setting mu_r=0 works. I recover my coefficients, and the model runs much faster, and in retrospect it is very obvious. It adds the mean of the receiver_fe to the sender_fe, but I don’t care about those values and there aren’t enough degrees of freedom to distinguish between individual sender and receiver effects if they’re unknown anyway. I was making it way more complicated in my head.
Thank you!
I’ll post the working simulation later today for posterity. It is obvious in retrospect, but maybe it’ll help someone else who got stuck in a similar logic loop.