Is multicollinearity a problem when fitting a regression model using ADVI?

Yes, we’ve seen many models that were “unacceptably slow” with NUTS but in 99% of the cases that slow-down is because of challenging posterior geometry and model reparameterizations and better priors made the model sample quickly.

I would try with Bambi which has some tricks to avoid common pitfalls like co-linearities.

1 Like