ENMeval Model evaluation

[Ione Arbilla]

Hello group,

I would like your opinion about the results of my MaxEnt models in Maxent.   I am thinking of choosing the model with the highest CBI.avg and lowest OR10, which would be the model highlighted in green. This way   I would have one threshold dependent and one threshold independent metric. My main concern is that the AUC values are quite low in comparison with the CBI values and I don't know what to think about that. Are CBI and OR10 enough to justify choosing a specific model?

The partition method that was used was checkerboard2 with 10000 bg points and 5311 occurrence points. I used rm higher than the default to avoid model overfitting. Also, models with Threshold features have been discarded because I understand that  hinge classes need to replace threshold features (Phillips, Anderson et al. 2017)and threshold features are very rare in an. I attached the table with model results so you can have a look at it.

Cheers,


Ione 

[Fabio David Alabar]

Hola GRUPO,

Tendrán alguna bibliografía referida a la formación de modelos de distribución en maxent con datos de ausencia??

Desde ya muchas gracias

[Jamie M. Kass]

Ione,

This is a bit late, but these are my two cents. This model may perform well via AUC, OR, and CBI, but it is also very complex (90 non-zero coefficients). You have a big dataset, and thus maybe a complex model is appropriate. Also, the P feature doesn't seem to contribute many parameters to the model, as LQH has 86 parameters. I would also take a look at the response curves to make sure they look ecologically realistic and compare the spatial prediction of this model to other similar ones and make sure it doesn't have any spatial patterns that look strange to you as the expert of this species.

Jamie

[Jamie M. Kass]

David, Maxent is a presence-background algorithm, and so should not use absence data at all. You should use a different algorithm, like a GLM, GAM, or even Random Forest or Boosted Regression Trees if you have absence data. Also, much absence data is unreliable and/or poor quality, so consider running a model with background data anyway at least to compare.

Jamie