[Biomod-commits] Predictors automatically dropped from models?

[Robin Engler]

Hi,

> I discovered that predictors are automatically dropped from GAM/GLM models,
> when their importance (via VarImportance) is zero. Is BIOMOD intended to
> work like that? Are there any possibilities to suppress the dropping?

Actually it is the opposite: the Variable importance is equal to zero
because the variable was not included/retained in the model. The
variable selection procedure depends on the particular model you use
(and the parameters you pass to it) but is generally based on whether
a variable explains a significant amount of variance in your data or
not.
BIOMOD computes variable importance after calibrating the models, and
variable importance does not and never will have any influence on the
selection of variables in the models.

If some of your variables are dropped with one modelling technique
(GAM), while being retained with another (GBM in that case), this
might be due to the fact that there is a high multicolinearity between
your variables.

Cheers,
Robin

********************************
Robin Engler
Spatial Ecology Group
University of Lausanne
Switzerland
********************************




On Mon, Aug 8, 2011 at 5:31 PM, Jonathan Heubes
<Jonathan.Heubes at senckenberg.de> wrote:
> Hi,
>
> I am currently working with R 2.13.0 and Biomod 1.1-6.9.
>
> I discovered that predictors are automatically dropped from GAM/GLM models,
> when their importance (via VarImportance) is zero. Is BIOMOD intended to
> work like that? Are there any possibilities to suppress the dropping? I am
> asking, because the evaluation of variable importance can be extremely
> different, pending on the models which are used. Example:
>
>> VarImportance
>
> $SP1
>
>      Var1 Var2 Var3 Var4
>
> GAM 0.000 0.973 0.000 0.000
>
> GBM 0.192 0.344 0.42 0.015
>
> The GAM model was calibrated with one predictor only (Var2)! All other
> variables were automatically dropped from the GAM, including Var3, which is
> suggested to be the most important one by GBM. When I used Models() and a
> priori excluded Var2, then Var3 became most important for GAMs. Actually
> this indicates collinearity, which is 0.77 for Var2 and Var3 (pearson
> correlation coefficient). This doesn’t occur with GBMs, because they are
> more robust against multi-collinearity? Any suggestions? Many thanks, cheers
> Jonathan
>
> PS: The settings in Models():
>
> Models(GAM = T, GBM = T, NbRepPA=1, strategy="random", nb.absences=1000,
>
>       NbRunEval = 5, DataSplit = 70, Yweights=NULL, Roc=TRUE,
> Optimized.Threshold.Roc=T,
>
>       Kappa=F, TSS=F, KeepPredIndependent = FALSE, VarImport=5)
>
> _______________________________________________
> Biomod-commits mailing list
> Biomod-commits at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/biomod-commits
>
>