[R-sig-eco] Reducing spatial autocorrelation

[Matthew Landis]

Dear list - Apologies if this is not a repost.  I sent it the first time 
in html format inadvertently and it hasn't shown up.

Corrado -
I meant to refer to a regression model - presumably you are going to 
build a regression model of sorts (although multivariate because of all 
the species) to see which of your explanatory variables is important.

For some more informtion:
Here is the website for SpaceMaker2, by Borcard and Legendre - they list 
a number of papers to describe the method.  
http://www.bio.umontreal.ca/casgrain/en/labo/spacemaker.html

I also found this paper by Stephane Dray to be helpful. 
http://pbil.univ-lyon1.fr/members/dray/articles/SD163.php

I must say, it is not all that simple to work out how to do the 
analysis.  But the results are very interesting.

M

Corrado wrote:
> Dear Matthew,
>
> thanks for your kind answer!
>
> The first approach you describe is the one I have been looking at until now.
>
> I am puzzled about the second one: I do not really understand it. What model 
> are you talking about, when you say "incorporate the spatial variation in the 
> model"? At the moment I have no model, just the data and I am trying to reduce 
> autocorrelation before analysing the data.
>
> Do you have any good reference (articles or books) about the approach you 
> mention?
>
> Thanks in advance
>
>
> On Wednesday 14 October 2009 13:11:04 Matthew Landis wrote:
>   
>> Corrado:
>>
>> The simplest way would be to take a subset of sites to maximize the
>> distance between them.  Say, choose 400 sites evenly spread over the
>> study area.  That would minimize autocorrelation to the greatest extent
>> possible, but you would be throwing away data.
>>
>> The second thing you could try would be to incorporate the spatial
>> variation in the model to control for it.  This way you can also study
>> the autocorrelation, see what spatial scales it is operating and what it
>> looks like and try to learn something from it.  Legendre, Borcard, Dray
>> and colleagues have developed some really interesting ways of dealing
>> with multivariate data and decomposing the variance into spatial
>> component vs. explanatory variables.  I believe it is called PCNM and
>> can be found in the spacemakeR package (don't think it is on CRAN - have
>> to do a google search).
>>
>> Good luck!
>>
>> Matthew Landis
>>
>> Corrado wrote:
>>     
>>> Dear friends,
>>>
>>> I have a large matrix of species (first 1100 columns) and environmental
>>> variables (last 36 columns) for approx 2000 sites.
>>>
>>> The distance between sites varies. Some sites are near to each other,
>>> others are far.
>>>
>>> I would like to select a subset of N sites (for example: 400 sites) with
>>> the minimum spatial autocorrelation. The aim is to obtain a significant
>>> number of sites to carry out some statistical analysis, but with spatial
>>> autocorrelation significantly reduced.
>>>
>>> Is there a procedure to do so in R? How would you approach the problem?
>>>
>>> The aim of the "reduction" is to then work on dissimilarities between
>>> sites that have the lowest possible spatial autocorrelation.
>>>
>>> Thanks
>>>       
>
>
>
>