Open
Description
As mentioned in #478, the convergence checks for additive models are probably still incomplete, even with PR #478 being merged now. I have added corresponding TODO
comments in the code, lines
projpred/R/divergence_minimizers.R
Lines 1022 to 1023 in 97c5bea
projpred/R/divergence_minimizers.R
Lines 1025 to 1028 in 97c5bea
@AlejandroCatalina, do you know if (and if yes, how) we could improve our check for convergence of the submodel fits from fit_gam_callback()
and fit_gamm_callback()
?
Metadata
Metadata
Assignees
Type
Projects
Milestone
Relationships
Development
No branches or pull requests
Activity
AlejandroCatalina commentedon Nov 23, 2023
Aside from the diagnostics
fit_s$gam
would provide I'm not sure how much we could inject it. I would suspect we can indeed run some diagnostics on top. ForGAM
I'd think that similar diagnostics toGLM
should work, as it's fitting an augmented GLM underneath, and forGAMM
one would need to check the quality of the laplace approximation, but this may not be very obvious from outside.fweber144 commentedon Nov 23, 2023
Actually, I was looking for some output element of the submodel fits returned by
fit_gam_callback()
andfit_gamm_callback()
that indicates convergence. So for GAMs, I would need approval thatprojpred/R/divergence_minimizers.R
Lines 1022 to 1023 in 97c5bea
gamm_fit_s$gam
are (wheregamm_fit_s
indicates an object returned byfit_gamm_callback()
). Implementing own diagnostics is not a bad idea, but usually, there is always a convergence indicator somewhere in the output object (so we wouldn't have to implement new diagnostics ourselves).AlejandroCatalinaF commentedon Nov 27, 2023
I believe that line is okay, but I haven't worked with that package for some time, so I'm afraid I can give you much insight regarding their internals.
fweber144 commentedon Nov 30, 2023
No problem, I understand. Then I guess someone needs to inspect this in detail in the future.