[R-sig-eco] R-sig-ecology Digest, Vol 71, Issue 13: SDM-model evaluation (via AUC or AIC)

[Claas Damken]

Hi Ludovico,

I have my doubts that AUC is suitable for count data analysis, since its based on the ROC, a measure of how well a parameter can distinguish between two diagnostic groups (i.e., presence vs. absence). But for count data, say 123 species at site a and 321 species at site b, most if not all model settings will fail to hit exactly those numbers. YOu could try and calculate the AUC for one of your models, but I would guess it would be lower than 0.5, but for the wrong reasons.

AICc and AICc weights are a better way to select model parameters, and Nagelkerkes N2 will give you additional information about the overal fit of your models. However, model selection via AIC or any of the other ICs is not straightfoward either and - FAR MORE IMPORTANT than choosing a statistical method - your data are "number of species per site" and not simply count data for one species? For the latter, an ecological parameter at a certain spatial scale might explain the number of individuals at site x very well, but if you are dealing with many species, is it ecologically speaking reasonable to assume that all species react to the same ecological parameters at the same spatial scale? Not sure what data you are working with, but for many systems, your results will be fairly predictable (not in a positive way).

Hope that helps

Cheers,

Claas

Claas Damken
Postdoctoral Fellow
Institute for Biodiversity and Environmental Research
Jalan Tungku Link, Gadong BE1410
Universiti Brunei Darussalam
Brunei Darussalam



________________________________________

   1. Re: SDM-model evaluation (Martin Weiser)
   2. Re: SDM-model evaluation (Frederico V. Faleiro)

Message: 1
Date: Fri, 21 Feb 2014 13:31:32 +0100
From: Martin Weiser <weiser2 at natur.cuni.cz>
To: Ludovico Frate <ludovicofrate at hotmail.it>
Cc: "r-sig-ecology at r-project.org" <r-sig-ecology at r-project.org>
Subject: Re: [R-sig-eco] SDM-model evaluation
Message-ID: <1392985892.22618.85.camel at zouzel.natur.cuni.cz>
Content-Type: text/plain; charset="UTF-8"

Hi,

I am not sure why AUC should be better than, say, AICc - second order
Akaike criterion and AIC weights (comfortably calculated by AICcModavg
package) or Nagelkerke's R2 in this case. Would anybody be so kind and
explain this for me?
I just want to make warning: do not put strong (or any) emphasis to
p-values of partial tests.

Best,
Martin

   Ludovico Frate p??e v ?t 20. 02. 2014 v 14:19 +0100:
> Hi all,I am trying to built a species distribution model with count data (Species Richness) using a GLM. My independent variable are computed at multiple-scales. In order to select which is the best scale for my model, I first calculate a bivariate model for each of the predictor for each scale. So, I know that one possible way to select the best predictor is the use of the AUC, but I don't know if is possible to use the AUC for count data. I have read about presence/absence data (for example in the package DISMO, function evaluate) but never for count data! Any suggestions?Thank you!
>
>
> Ludovico
> Frate
>
> PhD student (University of Molise - Italy)
> Environmetrics Lab
> http://www.distat.unimol.it/STAT/environmetrica/organico/collaboratori/ludovico-frate-1
> Department of Biosciences and Territory - DiBT
> Universit del Molise.
> Contrada Fonte
> Lappone,
> 86090 -  Pesche (IS)
> ITALIA.
> Cel: ++39
> 3333767557
> Fax: ++39 (0874) 404123
> E-mail ludovico.frate at unimol.it
> ludovicofrate at hotmail.it
>
>       [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-ecology mailing list
> R-sig-ecology at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology