Handling multiple cross-cutting (non-nested) groups in hierarchical model

[junpenglao]

@lwahedi

As you might expect, these intercepts are not stable: you could always add c to all sender effects and subtract it from all receiver effects. This makes the model particularly difficult to estimate, and gives me concerns about convergence overall.
If I were doing a fixed effects model, I would just exclude dummies for one of the categories, or set the value to 0 for that category. I am not sure how to set one value to 0 in PyMC3, or if there’s another options that would work well. Does anyone have any ideas?

When your model is over-parameterized like this, reducing the number of parameter is a good and straightforward way. You can try setting some mu_* to zero, for example, just do mu_r = 0

Also, you should avoid using pm.Flat as prior, a Normal distribution would be much better.