bioclim variables and Spatial Auto-Correlation

[Hira Fatima]

dear all!!

I am doing species distribution modeling using bioclim variables, however many studies suggested that these variables are highly HIGHLY correlated, so I did PCA on these variables.. and I would be using those PCs in my analysis, but how would I interpret those PCs in terms of environmental variables, studies suggest that it could be done by taking into account the PC loadings, but it aint workin for me... and since it is one of the objectives of my studies to look at the environmental predictors and how they contribute towards the model, I need to explain species distribution in terms of environmental variation.

i'll be very grateful if someone helps me out of this :)

Hira

[João Paulo Hoppe]

Hello Hira Fatima,

bioclimatic variables are really higly correlated, and PCA is the most common way to deal with this but not the only one. Please refer to Dormann et al 2012 (Ecography 35-1: pp. 1-20) to see a concise review on the subject.

What I do is reduce the numbers of variables, studying the PCs contribution and the loadings in said PCs, to choose a subset of these. Then, a Mantel test to see significance when correlated to geographical distance matrix.

Hope it helps you!

João Paulo

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/d/optout.

--

João Paulo Maires Hoppe
Bacharel em Ciências Biológicas
Mestrando em Biologia Animal
Links de interesse: Artigos, Lattes
UFES / Universidade Federal do Espírito Santo
Depto. de Ciências Biológicas
Av. Fernando Ferrari, s/n - Goiabeiras
Vitoria ES Brazil
29075-010

"There are no hopeless cases, only people without hope and expensive cases"

- Michael Soulé, Viable Populations for Conservation

[Hira Fatima]

Thankyou so much for your response.. but if i select variables even on the basis of highest loadings as shown in loadings chart (which i have attached here), some of the variables are still correlated...  i dont know how will i justify that...
Regards

[João Paulo Hoppe]

This is my personal opinion: as the bioclimatic layers are derived from a small set from the onset (tmax, tmin, precipitation), they will be correlated no matter what. We can reduce the multicollinearity by reducing the number of correlate layers, but some will remain. I think that the only way to use independent layers would be using directly the PCs from a PCA, but you'll have to justify that nevertheless.

On Fri, May 30, 2014 at 2:54 AM, Hira Fatima <fatim...@gmail.com> wrote:

Thankyou so much for your response.. but if i select variables even on the basis of highest loadings as shown in loadings chart (which i have attached here), some of the variables are still correlated...  i dont know how will i justify that...
Regards

--

You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/d/optout.

[Hira Fatima]

ok then i'll compare both the methods and see how the results are, how accurate the modeling is.. as you said i think the better way is to select variables on the basis of PC loadings as interpreting PCs seems difficult.. Thanks alot... :)

[Hira Fatima]

and you talked about the geographic distance matrix can you please elaborate it a little more..!!!

[João Paulo Hoppe]

Our reasoning is: if the environment is somehow related to the geographic distance, then a matrix dissimilarity correlation test (as Mantel) between the distance matrixes of environment and geographic data will be significant.
I do this in R, creating the distance matrixes with dist(), and doing the Mantel test with mantel(), available in the vegan package.

Doing this procedure tends to give us the same layers that a sequential regression (detailed in Dormann et al 2012) gives.

Right now, I'm writing a ms derived from my graduation research, and I intend to detail this method there.

On Sat, May 31, 2014 at 12:53 PM, Hira Fatima <fatim...@gmail.com> wrote:

and you talked about the geographic distance matrix can you please elaborate it a little more..!!!

--

You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent.
For more options, visit https://groups.google.com/d/optout.

[Hira Fatima]

Hello!
your idea sounds interesting and i am trying to dig deep into this mantel test and how it works for the variables... according to my understanding in sequential regression what we do is we take the most optimum variable (according to expert knowledge) and then we compare all the other variables with that first variable and then we calculate the contribution of all the dependent variables.. now coming back to this your suggestion, since we know environment is related geographically, we take the geographic distance and then compare all the environmental distances with that taking into account the significant ones.... is this your idea??? and we can calculate geographic distance using the point data but how would you do that for spatial environmental datasets?
also goodluck with your work :)