Open nehill197 opened 3 months ago
Started working on this and was able to validate the deviance calculation by comparing to the output of a VGAM::vglm()
model.
We need to figure out if each row modelled by a multinomial should have a single deviance value or one for each column in that row.
This may require an "extended" way to calculate the sign for the deviance residuals (i.e., it isn't as simple as taking the sign of a single value). There is a method to do this detailed here (https://convegni.unica.it/cladag2015/files/2015/10/Romeo1.pdf), however I find it unclear.
This might also be worth a read:
See if this post has any validity https://stats.stackexchange.com/questions/186560/what-is-multinomial-deviance-in-the-glmnet-package