Hi,

Why not follow Hirzel approach (paper link above)? They compare Boyce index with k-fold median value.

https://cran.r-project.org/web/packages/ecospat/ecospat.pdf

http://rpackages.ianhowson.com/cran/ecospat/man/ecospat.boyce.html

http://www2.unil.ch/biomapper/Download/Hirzel-EcoMod-2006.pdf

Thanks Paulo.

Hi Lorenzo,

I agree with Paulo--the Continuous Boyce Index is probably better than the normal Boyce Index.  FYI, I created R code to calculate the CBI (http://www.earthskysea.org/r-code/). 

And yes, it does make sense to use k-folding with this test statistic and with most others.  However, if you're using CBI to weight individual models to create an ensemble, be aware that CBI can be negative (meaning worse than random), so these models should not be included in an ensemble.  Better yet, if you're using the same algorithm, compute a series of models with k-fold splits and calculate the test statistics.  Then, use all of the presence sites to calculate an "all-points" model.  The latter will most likely be as accurate or even more than the individual k-fold models because it uses more data.

By the way, do you mean you're using the BIOCLIM model?  It's very easy to interpret and conceptualize, but it's often one of the worst-performing models.  You might try some of the other models available in dismo, especially Maxent or generalized additive models.

The Continuous Boyce Index - if critically analyzed-can show you the route