[R-sig-Geo] spatial regression for large data sets

[Seth J Myers]

Hi everyone,

I'm trying to model a binary response using logistic regression for a large data set with spatial autocorrelation issues.  The mixed models in SAS and R that can include spatially correlated errors cannot handle the large NxN matrix needed for their methods.  Past around 700 meters, spatial autocorrelation is negligible.  So, it seems that a sparse matrix with zeros for pairs of observations separated by > 700 meters would help.  I've been searching and not found a way to implement this in R using established packages.  Any advice on methods to try for a large data set with spatial autocorrelation is welcome.  I'm currently reading to see if generalized estimating equations or some of the detrended krigging fuctions may help.  If I sample systematically with a 700 m distance, I will lose substantial information and so would like to avoid this. Thanks, Seth