[R-sig-Geo] Spatial autocorrelation in GLM/GAM/CART binomial family

Arnaud Mosnier a.mosnier at gmail.com
Fri Apr 29 15:10:16 CEST 2011


Federico,

I recommend reading the following paper :

Methods to account for spatial autocorrelation in the analysis of
species distributional data: a review
Carsten F. Dormann, Jana M. McPherson, Miguel B. Araújo, Roger Bivand,
Janine Bolliger, Gudrun Carl, Richard G. Davies, Alexandre Hirzel,
Walter Jetz, W. Daniel Kissling, Ingolf Kühn, Ralf Ohlemüller, Pedro
R. Peres-Neto, Björn Reineking, Boris Schröder, Frank M. Schurr,
Robert Wilson
Ecography Volume 30, Issue 5, pages 609–628, October 2007

Hope this help,

Arnaud


Date: Fri, 29 Apr 2011 04:14:44 +0000
From: Federico Tomasetto <federicotomasetto at hotmail.it>
To: <r-sig-geo at r-project.org>
Subject: [R-sig-Geo] Spatial autocorrelation in GLM/GAM/CART binomial
       family
Message-ID: <DUB107-w4222F08EB0EE24BDE52719C89A0 at phx.gbl>
Content-Type: text/plain


Hi List,


I have 1227 data plots regularly sampled in a grid
of 1 x 1 Km in which species richness and abundance of 682 vascular plant
species have been recorded.

I constructed also a database of environmental
variables with GIS. My first goal was to compute a GLS model to explain species
richness with the environment. To do so I used GLS + spatial structure as it is
explained in Pinheiro-Bates, Mixed-Effects Models in S and S-PLUS (pag
260-267). The model spatial vs non spatial have worked ok, showing lower AIC
values in GLS + spatial model than GLS + no spatial. So my data is not
spatially independent and show spatial autocorrelation.

Now, I intend to compute GLM/GAM/CART models with
binomial data family to compute the abundance of species in each plot and the
environment taking into account the spatial autocorrelation. The final
objective would be to compare those models (spatial and non spatial and check
which one is the "best").

I read that a way to do a spatial GLM is using the
glmmPQL function of MASS and putting all the observations in the same group for
the random effect. If this is the way to do it,
how about GAM and CART?

Do you have any idea how to deal with 682 species and model
fitting and if there is any other methods to do what I want (comparison of
spatial and non-spatial)?
Many thanks for any kind of help
Federico



More information about the R-sig-Geo mailing list