[R-sig-Geo] question about fitted values form geoR - results 'too good'

[Johan Van de Wauw]

Kriging is an exact interpolator. It will always reproduce the values
that you use to predict if they are located on the same locations. You
should use an independant data set (or some sort of cross-validation)
to check how good your prediction is.

On Nov 8, 2007 10:47 PM, Ken Nussear <knussear at usgs.gov> wrote:
> Hi
>
> I'm using geoR for some spatial linear models and I'm getting
> surprisingly optimistic values from the spatial models relative to the
> non-spatial, even when the models appear to be performing about
> equally (by AIC comparison)
>
> For example
>
> This model relating encounter rates of lizards to a soil substrate
> parameter gives
>
> > > summary(m2)
> > Summary of the parameter estimation
> > -----------------------------------
> > Estimation method: maximum likelihood
> >
> > Parameters of the mean component (trend):
> >  beta0  beta1
> > 0.0312 0.0024
> >
> > Parameters of the spatial component:
> >    correlation function: exponential
> >       (estimated) variance parameter sigmasq (partial sill) =  0.0082
> >       (estimated) cor. fct. parameter phi (range parameter)  =  797.1
> >    anisotropy parameters:
> >       (fixed) anisotropy angle = 0  ( 0 degrees )
> >       (fixed) anisotropy ratio = 1
> >
> > Parameter of the error component:
> >       (estimated) nugget =  0.002
> >
> > Transformation parameter:
> >       (fixed) Box-Cox parameter = 1 (no transformation)
> >
> > Maximised Likelihood:
> >    log.L n.params      AIC      BIC
> >  "53.44"      "5" "-96.87" "-86.57"
> >
> > non spatial model:
> >    log.L n.params      AIC      BIC
> >  "51.99"      "3" "-97.98"  "-91.8"
>
>
> With a difference in AIC of only about 1.
>
> However looking at the predicted values versus the fits for the model
> The spatial model fitted values appear to be some how too good.
>
>  > cor(fitted.likGRF(m2, spatial=TRUE), td$Crotaphytus)
> [1] 0.9934701
>
>
>  > cor(fitted.likGRF(m2, spatial=FALSE), td$Crotaphytus)
> [1] 0.2522837
>
> So I don't get how the spatial model with only a delta AIC of 1 can
> have a correlation with the dependent variable that is this high. Am I
> mis-interpreting the values I'm getting from the fitted call, or is
> something amis.  I've tried this with different data sets and I'm
> getting the same result.
>
> Thanks
>
> Ken
>
> _______________________________________________
> R-sig-Geo mailing list
> R-sig-Geo at stat.math.ethz.ch
> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>