I found R to do this analysis!!

Always better to do a correlation analysis for your variables prior to running a model because if some of your variables are correlated, it will bias your model outcome.  If variables are significantly correlated, you will need to select one and discard the other.  In any case, the fewer variables in the model the better, as otherwise you will get a problem of over-fitting the data.

Finally I have done my corr analysis.. Thank you!

Hi Hossein, I wrote you by inbox. If was useful, just let me know.

Even if used a correlation analysis its bit tricky to select one variable over the other. Say if A and B correlated what you exclude, A or B?  For that basic understanding of the ecology of the modeled organism is needed, isn't it?

For a range of procedures for variable selection in SDM you may consult the attached and its preceding articles.

Article Transferability of species distribution models: The case of ...

This paper covers the topic in detail

Dormann CF, et al. (2012) Correlation and process in species distribution models: Bridging a dichotomy. J Biogeogr 39(12):2119–2131.

Sure. It is not a only methodological issue; it is a conceptual approach to make robust the result. It is better to find variables linked with a hypothesis to based the analysis. You can use decisison or correlation trees. 

Dear Andrea,

please can you send me to the procedure involved for correlation analysis,  as i am also not knowing how to extract the data from bioclimatic variables to run a pca..

thank you

Dear Andrea,

I am in need for running an SDM for tree species. I would like to ask you if you have explanation on how to run the correlation analysis in R ? What modules or packages did you use ?

Thank you very much for your help :)

Evan

Dear Andrea,

I prefer doing pairwise (Pearson/Spearman/Kendall) correlation and multicollinearity (Condition Number, CN, Variance Inflation Factor, VIF) analysis before building the species distribution model (SDM). If you know R statistical software, you might find interesting its package called 'corrplot' that enables visualizing the pairwise correlation matrix of your data. I think these maximum values may serve as thresholds of your correlation measures: pairwise correlation - 0.7 to 0.8; CN - 30; VIF - 10 to 50. [One of my papers is attached es an example in which we used these thresholds.]

If your predictor set is facing with correlation/multicollinearity problem, you have three possibilities to deal with it. (1) You may eliminate predictors one by one (in a stepwise manner), which needs expert decision and interative correlation analysis. (2) You can aggregate predictors and create non-correlating new predictors (principal components of Principal Component Analysis (PCA)). (3) Some SDM methods can deal with multicollinearity in the training predictor database, however you may pay attention on predicting to future time period or far locations where the correlation structure of the dataset is outside those structures that were learned by the model. In such case, you will extrapolate, which cause high uncertainty of your predictions.

Please find attached a paper about multicollinearity and how SDMs can face with them. Some links to 'corrplot' package are also attached.

HTH,

Ákos

https://cran.r-project.org/web/packages/corrplot/vignettes/corrplot-intro.html

https://cran.r-project.org/web/packages/corrplot/

Article Collinearity: A review of methods to deal with it and a simu...

Article Implementation and application of multiple potential natural...

You will find a lots of paper on both side, do and do not do correlation test on variables. 

If you are using conventional statistics, you should do it. and if you are doing machine learning things, you may not need to do it. 

Indeed Martin, an a priori decision on the the removal of one of two correlated variables is arbitrary and often not transparent. The issue exacerbates, as you imply for larger numbers of variables.

For a transparent and reproducible (i.e. scientific) way to remove redundant variables in maxent SDM, please our endemics and bear articles.

regards

@Martin,

Did you try PCA? This method lumps together highly correlated variables and if the structure of variables permits, produces a few "Principel components" that can explain a considerably high share of variation in your variables. These components (or axes) can be used as "composite variables" which can be attributes to the prominent variables that correlate with the corresponding axes highly. Besides, these PCA axes are orthogonal, that is, correlation among them is exactly 0. This permits using multivariate regression (otherwise this method suffers strongly from collinearity -- when there is significant correlation among the variables).

You can simply do a test run using all variables. Then do a correlation and find the correlated variables. Use the percentage contribution (results from the test run) of each variables to select the most contributed uncorrelated variables to build the SDM.

@Nayana, maybe you mean a clustering of all variables? that's what I did

You also can run a test using VIF package in R to identify most correlated variables.

https://cran.r-project.org/web/packages/VIF/index.html

Thank you!. The package Hmisc is useful for this analysis too.

Yes this is required, but you have to crosscheck this selection.

Sdm tool in Arc GIS or ENM tool are also useful

http://www.danwarren.net/enmtools/builds/Release/ENMTools_User_Manual%201.1.pdf

https://github.com/danlwarren/ENMTools

Available research suggests both

Not only correlation/multicollinearity and the ecological interpretability (as mentioned by Martin E. Timana ) are important when you select from the available bioclimatic variables, but the way how the month/quarter the bioclimatic variable is based on will shift in time (and in space). Please refer to this paper:

Article The way bioclimatic variables are calculated has impact on p...