Dear all,
I modeled the distribution range of species of interest with presence-only points using biomod2 package within an ensemble framework and made a projection to other space that I have some independent presence points of species. I would make a statistical test to assess the accuracy of the model predictions in the projected region. I saw that some authors (e.g. Thuiller et al. 2005: Niche-based modeling as a tool … in Global Change Biology 11: 2234-2250) used chi-square tests to assess model performance as absence records were not available, but I am afraid that for doing chi-square test it needs to classify our prediction values into two presence and absence categories based on the specific threshold in which this is an another matter as well. I wonder if you help me on this matter.
Yes ... check out the "Model Evaluation" chapter of our 2011 book, for an introduction to the topic. Another good review is Fielding and Bell 1997, which you can grab at http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.463.359&rep=rep1&type=pdf.
But more specifically, you can use threshold-dependent (yes, this requires a threshold, which should be a least presence threshold, perhaps modified to take into account the existence of some error in your input data) tests, such as chi square, although a cumulative binomial test is better and simpler.
More popular are ROC approaches. The typical ROC test is not appropriate for ecological niche modeling applications -- see Lobo et al. paper attached. However, a partial ROC modification can work well ... see Peterson et al. 2008 attached. An easy implementation is available at http://shiny.conabio.gob.mx:3838/nichetoolb2/.
Hope this helps. ATP
Article Rethinking receiver operating characteristic analysis applic...
Article AUC: a misleading measure of the predictive distribution mod...
Article Ecological Niches and Geographic Distributions by A. Townsen...
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS