brglm v.s. brglm2 brglmFit did not converge about brglm2 HOT 6 CLOSED

Yes. For binomial-response models brglmFit with type = AS_mean produces the same estimates as brglm does. So, regardless of what you use the estimates should be finite. Can you send me a reproducible example?

from brglm2.

I use my data and cannot reproduce the situation again. I think the warning message occurred during the bootstrap steps. Let me try if it can be reproduced again.

from brglm2.

Hi Joy, please note that for some reason whenever you are writing all past discussion is posted too (perhaps you use you mail client?). So please use the GitHub interface for the discussion, as I would like to keep the issue discussion clean for the benefit of other users.

Having said that I cannot find the attached code and hence cannot replicate the issue.

Please repost.

from brglm2.

Dear Dr. Kosmidis,

I'm sorry. I did not know that when I write mail, it will be posted.

Attached are the data and code I used. I think the warning message "brglmFit: algorithm did not converge" came out during the bootstraps, By the way, I also used another package called "medflex" in my works.

test for brglm2.zip

from brglm2.

Hi Joy,

Thanks for the reproducible example!

neModel seems to be doing bootstrap by case resampling or a associated variant and your data set has covariates that are highly imbalanced. For example, the dummy corresponding toStage == "not reported" has 1002 zeros and 12 ones. So, the bootstrap datasets can end up having very few ones or only zeros with positive probability.

So, I think it is not brglmFit's fault. brglmFit now tries hard to find good starting values and fit the model despite the highly imbalanced datasets, and it actually does this very well and better than brglm and glm! It only fails in a handful of cases of the above kind, and both brglm::brglm.fit and glm.fit fail to converge on those and more cases.

Overall, I think neModel can be either given an option to ignore bootstrap samples where fitting has not been possible (returning a warning potentially), or use residual-resampling bootstrap if the methodology permits. But this is the call of neModel's developers. Perhaps you can ask them directly?

I hope that helps. Please let me know if you find any bugs with brglmFit.

All the best,

Ioannis

from brglm2.

Dear Dr. Kosmidis,
Thank you for your kind assistance and good advice!

from brglm2.