Open svteichman opened 3 months ago
@svteichman I'm just chatting w David and he confirmed he had no issues with brglm2
. I was thinking of logistf
, which we had significant issues with when implementing happi
.
David did say brglm2
breaks/takes ages for multinomial logistic regression with many categories (large $J$), such as we see in radEmu-type problems.
I think a next step after #15 is to implement a Firth penalty for multinomial as well. From there I think we will mostly have what we need to share this with the collaborator who requested it 🥳
Look into stability of
brglm2
, then add as an option (replacingglm.fit
withbrglm.fit
as needed) for running glm to adding Firth penalty functionality. This will require a bit of craftiness to go from a design matrix (whatraoBust
uses to drop a column to estimate under the null) to a formula and data object (whatbrglm
expects).