[R] Help with Krige.conv using linear models

[Jennifer Lenz]

A majority of my data makes a kriged map perfectly using an exponential 
model for the semivariogram to fit my data and then going through the 
commands variofit() to define the model and then krige.conv() to use the 
model to predict values in a grid.  But…one set of my data appears to be 
linearly correlated for the first 5000 meters and not correlated beyond 
that. I have been having problems using krige.conv() to get a decent 
kriged map using the linear model.  The code I am using from my data is 
as follows:

 >modeltest=variofit(variotest, weights=’cressie’, cov.model=’linear’, 
ini.cov.pars=c(80,1))

The output parameters are tausq = 9.855, sigmasq = 0.0087, phi=1.0

 >krig=krige.conv(data, krige=krige.control(type.krig=’ok’, 
obj.model=modeltest), locations=pred.grid)

At this point, krig$predict values have little to no variability (1.897 
+/- 0.004), where I would expect values between 0 – 15.

By running the same data, except using an exponential model such as:

 >modeltest=variofit(variotest, weights=’cressie’, cov.model=’exponential’)

This model appears to fit the data (only the first 5 km) about the same 
as the linear model except now the output parameters are tausq=10.02, 
sigmasq=318909, and phi = 3714567.  And calling krige.conv() again, the 
predicted values are in the range that I would expect, and the kriged 
map looks fine.

I’m not sure if this has something to do with how the kriging is using 
the model beyond the 5 km.  It seems like I need to be able to set the 
function to only apply the linear weighting to the linearly correlated 
portion (data less than 5km away), and the rest to 0, but I’m not sure 
how to do that (or maybe I’m completely going down the wrong track).

I’m using Windows XP pro OS with R 2.1.1.  Please help the neophyte 
statistician.  To figure this out is the only Christmas gift that I 
need.  Thanks.