Request new feature for Firth’s logistic regression

[Wallace Wang]

Hi Developers,

Can I request a new feature for Firth’s logistic regression? As the GLM logistic regression has no power for (qusi-)separation data by the wald test. Firth’s logistic regression is good to compensate for this. Ref: A Solution to Separation in Binary Response Models. Firth logistic regression got great attention in recent genetic fields (Ref: Firth logistic regression for rare variant association tests).

Best regards
Wallace

[Milan Bouchet-Valat]

You're welcome to request, but this will only be implemented if somebody is willing to devote time to it. Also maybe it would better live in a separate package (not sure -- I guess it depends on whether it can benefit from reusing internal code from GLM.jl or not).

[José Bayoán Santiago Calderón]

I would agree that GLM scope is limited to IRLS distribution/link regressions. More involved models are likely to be developed and hosted at a different package (e.g., MixedModels.jl). Econometrics.jl has in its dev plans more advanced probability models... However, I would love to have a lightweight package for "safe" categorical probability models. Most of the ecosystem would likely benefit from it. The reference you mentioned is one approach. I believe the standard is to drop features/rows until it is no longer completely separable (e.g., Stata).