[R-sig-Geo] comparing non-spatial and spatial generalized linear models

[Beth Crase]

Hello Elaine,

It depends on the focus of your study. 
(1)If you want to work out which of your three explanatory variables is the 
most important for bird richness, then you will probably compare parameter 
estimates from non-spatial and spatial (autoregressive) models. See Dormann et 
al (2007) Ecography 30:609-628 for excellent examples and code. They use 
simulated data, and you use actual data so you will not know the true value of 
your parameters. If there is a lot of autocorrelation in your data then your 
parameter estimates will be poor. The effect of autocorrelation in your data 
is to inflate the importance of variables - but you may not know which ones or 
by how much. 

(2)If you want to make good predictions, then compare the predicted values 
from your spatial and non-spatial models to your actual observed values. For 
this you could calculate AUC (or ROC) and percent of deviance explained by 
partitioning your data and using some for training, and some for testing. 
Cross validation would be better. And an independent data set, better yet. See 
Betts et al (2006) Ecological Modelling 191, 197-224.

(3)If you just want to show that there IS autocorrelation in your data, then 
you could calculate or plot Moran's I, and show any differences in parameter 
estimates from the non-spatial and spatial models.  
Cheers
Beth.