Support new ICAR distribution

[lemartinet]

Hi Tomas!
I was wondering if the new ICAR model that was added in pymc 5.8 could be potentially supported in Bambi, or if the current architecture would not be able to handle this.
Thanks for your thoughts!
Manu

[tomicapretto]

Hi Manu! Thanks for opening the issue.

I don't think it's impossible but:

• I see pm.ICAR asks for an adjacency matrix. What should be the syntax to specify that in Bambi? As columns of a dataframe it could be very tedious and unpractical.
• How should we treat the centered/non-centered parametrization? I think Bambi should automatically handle this under the hood. But, how do we do it? I guess we need a wrapper on top of the PyMC distribution, or some kind of special treatment, to automatically get mu out of phi and sigma depending on the case. I think it would be the first time we do something like this.

I can't promise anything but I'll be thinking about this.

[lemartinet]

Thanks for sharing your thoughts on that, it's helpful to understand what's needed under the hood!