Skip to content

autologistic modelling in R

5 messages · Charlotte Bell, milton ruser, David Katz +1 more

Charlotte Bell
# 
Hi,

I have spatially autocorrelated data (with a binary response variable and
continuous predictor variables). I believe I need to do an autologistic
model, does anyone know a method for doing this in R?

Many thanks

C Bell
Roger Bivand
# 
Charlotte Bell <charlotte.bell <at> sheffield.ac.uk> writes:
There are several approaches that you could try. One direct spatial approach is
the off-CRAN Rcitrus package:

http://www.leg.ufpr.br/Rcitrus/

which although the documentation is in Portuguese, should get you most of the 
way there. You could also look at geoRglm on CRAN, which handles a similar 
setting in a geostatistical way. You may also find it helpful to look at 
the handling of spatial autocorrelation in the nlme package in a GLMM context, 
using the CorSpatial approach. If you like, you could also look at a GAMM 
approach in mgcv. The glmmBUGS package can be used for preparing a GLMM for 
running in *BUGS if the spatial autocorrelation is expressed through a 
spatial weights matrix rather than as a function of distance.

Hope this helps,

Roger Bivand.

PS. RSiteSearch on autologistic does find:

http://finzi.psych.upenn.edu/R/Rhelp02a/archive/147538.html

which is a posting by Elias Krainski on R-sig-geo, where a further link
is given for a forthcoming stLattice package.
1 day later
David Katz
# 
Curiosity and Google lead me to this paper which may be of interest:

Assessing the validity of autologistic regression
Purchase the full-text article



References and further reading may be available for this article. To view
references and further reading you must purchase this article.

Carsten F. DormannCorresponding Author Contact Information, a, E-mail The
Corresponding Author

aDepartment of Computational Landscape Ecology, UFZ Centre for Environmental
Research, Permoserstr. 15, 04318 Leipzig, Germany

Received 11 July 2006; 
revised 30 April 2007; 
accepted 7 May 2007. 
Available online 20 June 2007.

Abstract

In autologistic regression models employed in the analysis of species?
spatial distributions, an additional explanatory variable, the
autocovariate, is used to correct the effect of spatial autocorrelation. The
values of the autocovariate depend on the values of the response variable in
the neighbourhood. While this approach has been widely used over the last
ten years in biogeographical analyses, it has not been assessed for its
validity and performance against artificial simulation data with known
properties. I here present such an assessment, varying the range and
strength of spatial autocorrelation in the data as well as the prevalence of
the focal species. Autologistic regression models consistently underestimate
the effect of the environmental variable in the model and give biased
estimates compared to a non-spatial logistic regression. A comparison with
other methods available for the correction of spatial autocorrelation shows
that autologistic regression is more biased and less reliable and hence
should be used only in concert with other reference methods.
charlotte.bell wrote:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Roger Bivand
    

    

      
      #
    

  


  

      
David Katz <david <at> davidkatzconsulting.com> writes:

            

      
      
      
    

        
...

            

      
      
      
    

        
Good point and reference. My understanding is that the term "autologistic" is 
used both of a GLMM with a spatially structured random effect, and of a GLM 
with the product of the response and a matrix of spatial weights included on 
the right hand side as an "autocovariate". 

I believe that the article finds that the use of such an "autocovariate" can 
lead to the problems described, for example when the "autocovariate" is 
strongly correlated with the other right hand side variables. A properly 
specified GLMM should not suffer from the same problems.

Roger Bivand