[R-sig-eco] (no subject)

[Sven Adler (2)]

Dear Chris, Dear Edzer,
in mgcv you can include a correlation structure like correlation=corExp(form=~X+Y) as corGaus works too (you have to use the gamm function). I have compared for several data sets the performance of kirgging and gam and so far I hav't found
significant differences between the redisuals of both methods (mostly marine data). Problematic is the gamm with large data sets, but you can restrict the calculation of the correlation matrix (Wood 2006), that not all data will be include (only those that are colse to each point as in krigging too) Modellint species response this is important as different hot spots will disturbe your correlogram.
Using s(X,Y) works in mgcv fine, but in my oppinion the coordinates should only use as predictors if I have no other environmental variables as plants and animals choose thouse places where they can founs somthing to eat, it is warm the salinity has good values and so on.....
 "I never
understood why people would test for spatial autocorrelation, as
non-significant autocorrelation will only indicate that the sample size
is small."

 Maby it is here the wrong place (r-sig-geo) but it is an important question for ecological research too.
If I found no improvement when including spatial autocorrelation in the model, following the principle of simplicity,  should I remove it from the model?

Sven