[Biomod-commits] reporting VarImportance

[Viorel Popescu]

Dear Wilfried,

Thank you for your advice...

> This is a very good point. I also agree with you there are not enough
papers reporting the importance of the variables.
> Make sure you do not use variables that are highly correlated BTW, else,
it will not mean too much for those variables. This is important.

I did select the initial variables using a 0.7 correlation coefficient
threshold, so none of the variables were highly correlated.

> Barplot should indeed work perfectly fine. You may, at the same time,
look at whether some groups of species have similar variable importance.
> If you have run many repetitions and PA selections (if any), you may also
consider reporting (in Supp Mat) the bar plots for each models to depict
whether some models are highly unstable on those aspects and some are not.

I have tried both barplots with SE, as well as boxplots, and found that the
boxplots are much more informative of the spread of values because the data
was highly skewed.

> I guess it depends on the questions. I am afraid I am quite uncomfortable
about negative correlations in this aspect. What does it really mean? note
that we do not return any statistical test. Maybe, in those cases, a
statistical test > will show they are not different than 1. My take on that
is that either you let the 'true' value (usually, negative values are
anyway close to 0) or you bound the correlations between 0 and 1 under the
hypothesis that you are only
> interested into positive correlation, and than negative, for you, means
no correlation.

I ended up bounding the correlations between 0 and 1, as negative
correlations (VarImportance values >1) were very few and made up a very
small fraction of all the values.

Cheers,
Viorel


On Thu, Mar 28, 2013 at 11:04 PM, Wilfried Thuiller <
wilfried.thuiller at ujf-grenoble.fr> wrote:

> Dear Viorel,
>
> This is a very good point. I also agree with you there are not enough
> papers reporting the importance of the variables.
> Make sure you do not use variables that are highly correlated BTW, else,
> it will not mean too much for those variables. This is important.
>
> > I was thinking about averaging the VarImportance scores across all my
> > species and models, and present them using a bar graph with SE bars. Is
> it
> > OK to do this across models, or does it need to be done for each model
> > separately?
>
> Barplot should indeed work perfectly fine. You may, at the same time, look
> at whether some groups of species have similar variable importance.
> If you have run many repetitions and PA selections (if any), you may also
> consider reporting (in Supp Mat) the bar plots for each models to depict
> whether some models are highly unstable on those aspects and some are not.
>
> > Also, how do I deal with the values >1, which denote negative
> > correlations between the original predictions and the ones with the
> > permuted variable?
>
> I guess it depends on the questions. I am afraid I am quite uncomfortable
> about negative correlations in this aspect. What does it really mean? note
> that we do not return any statistical test. Maybe, in those cases, a
> statistical test will show they are not different than 1. My take on that
> is that either you let the 'true' value (usually, negative values are
> anyway close to 0) or you bound the correlations between 0 and 1 under the
> hypothesis that you are only interested into positive correlation, and than
> negative, for you, means no correlation.
> You may also consider than the ranking is importance more than the "true"
> value of the estimate.
>
> Hope it helps,
>
> Wilfried
>
>
>
> >
> > Thank you in advance for any advice...
> >
> > Ceers,
> > Viorel
> > _______________________________________________
> > 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
>
> -----------------------------
> Dr. Wilfried Thuiller
> Laboratoire d'Ecologie Alpine, UMR CNRS 5553
> Université Joseph Fourier
> BP53, 38041 Grenoble cedex 9, France
> tel: +33 (0)4 76 51 44 97
> fax: +33 (0)4 76 51 42 79
>
> Email: wilfried.thuiller at ujf-grenoble.fr
> Personal website: http://www.will.chez-alice.fr
> Team website: http://www-leca.ujf-grenoble.fr/equipes/emabio.htm
>
> ERC Starting Grant TEEMBIO project:
> http://www.will.chez-alice.fr/Research.html
>
>


-- 
Viorel D. Popescu
David H. Smith Conservation Research Fellow
University of California - Santa Cruz &
Simon Fraser University, Biological Sciences
(604) 340 4228
https://sites.google.com/site/vioreldpopescu/