@Michielstock Ha yes I agree with that of course - binomial or multinomial regression often requires regularisation. But a plain L2 penalty/independent Gaussian prior as in scikitlearn works less well than Firth's penalty (Jeffrey's prior) though, cf https
Here's a comparison to ridge regression. https://t.co/50x6xBoKSR
@davidshor They are conceptually similar I’d say. But Firth is the almost exact penalty needed to give unbiased coefficients. Here’s a paper by @Georg__Heinze and others. Pp. 2-3 explain the differences and similarity nicely, I think. https://t.co/m2mJWa