[Biomod-commits] Convergence problem with GLM

[Damien Georges]

Hi Frank,

For the way non-converging GLM are handle within biomod, you are right :
- only a warning is printed
- they are included in ensemble models

For the question of whether they should be keep or not... I don't really 
know.. but I would say that we can keep them if they are "good enough" 
(SRE models return also binary prediction and they are kept for 
ensemble-models building). Moreover, putting a threshold for  models 
selection at ensemble-modelling step should prevent you from keeping 
unrealistic models.

I'm also interested in what people think about it.

Waiting for other comments,

All the best,
Damien


On 14/06/2013 16:26, frank.breiner at wsl.ch wrote:
> Hi all,
>
> when having only few presences and the ratio of the number of predictors
> to the number of presences is far from the recommanded 1/10 rule, GLM
> models often do not converge. This could result in binary predictions as
> shown in the attached figures.
> How is the problem handled in Biomod? I think there is only a warning that
>
> the Model did not converge but the non-convergence models are also
> included in the ensemble. Correct?
>
> As you can see from the figures the non-convergence GLM (top left) output
> fits with the predictions of GBM (top right) and Maxent.(bottom left).
> Bottom right is a self made Ensemble including also GLM.
>
>
>
> For me it would make sense to include GLM output in the ensemble as the
> trend is the same. Does Biomod keep non-convergence models for the
> ensemble? What do you think? Keep it or remove it?
>
> By the way: AUC values for the shown models suggest that GLM is not a bad
> model:
>
> GLM:            0.981
> GBM:            0.998
> MAXENT: 0.997
> Ensemble:       0.998
>
> Looking forward to your responses.
>
> Best,
> Frank
>
> Frank Breiner
> Eidg. Forschungsanstalt WSL
> Biodiversität & Naturschutzbiologie
> Zürcherstrasse 111
> CH-8903 Birmensdorf
>
> Tel.: +41 44 7392 418
> Fax: +41 44 7392 215
> http://www.wsl.ch/info/mitarbeitende/breiner/index_DE
>
>
> _______________________________________________
> Biomod-commits mailing list
> Biomod-commits at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/biomod-commits