AUC values and optimal model choice

[Janelle Deshais]

Hi All,
I am a graduate student modeling plant disease hosts throughout
Redwood National Parks, California. I am hoping for clarification on
how to choose my optimal model. I was disheartened by my AUC values of
0.60 and 0.67 for my two models, until I reread Phillips et al.
(2006), which states "If the species' distribution covers a fraction
"a" of the pixels, then the maximum achievable AUC can be shown to be
exactly 1 - a/2. Unfortunately, we typically do not know the value of
a, so we cannot say how close to optimal a given AUC value is"

Am I correct in deducing that if my tree species (tanoak) truly does
exist in 70% of my study site, then my largest AUC value would be 65
(1-0.70/2)?

Best,
Janelle
--
Janelle Deshais
Graduate Student
Humboldt State University
Dept. of Forestry and Watershed Management
1 Harpst Street, Arcata, California 95521-8299

[Bill Peterman]

Hi Janelle--

You are correct that if your species is more or less ubiquitous across
your study area that you will always get a relatively low AUC. Rather
than put such a dependence on a high AUC value, you can generate null
models to get a distribution of AUC values and then assess whether or
not your model AUC is significant compared to the null.

There is a null model approach that has been detailed by Niels Raes
and Hans ter Steege (http://www3.interscience.wiley.com/journal/
117966250/abstract?CRETRY=1&SRETRY=0). It cannot be directly
implemented in Maxent, but has been adapted to use Maxent outputs and
framework for running the random null models. This has also been
discussed previously on the forum. If you search around a bit, you can
find more.

Good Luck--Bill