I'm trying to work with bambi models for Poisson GLMs with relatively low frequency (insurance claims data). The priors that are automatically generated for me tend to have rather large sigmas (can get to the order of 10s). Because the parameters get exponentiated, this implies very large frequencies (sigma = 50, at 1-sigma implies 5e21 at the exponentiated scale). This invariably leads to the ValueError: lam value too large, even with purely prior predictive checks.

Currently I deal with this by setting by hand all of the priors, but it would be nice to adapt the automatic prior selection to take care of the link function. Either that, or have the ability to specify a maximum to sigma that would apply to all the priors.

I scanned through the issues and couldn't find anything similar to this, so apologies if this is a duplicate.

Thanks!
Omri