Re: Errors in prediction

[Edzer Pebesma <edzer@pangea.Stanford.EDU>]

The `V' field (location-specific measurement errors) works for
data for which the variogram of measurement error-free values
were used to model the variogram.

Non-constant measurement errors implies that your data are
non-stationary (w.r.t. variance). If there is little variation
in the measurement errors, you can model their average behaviour
with a nugget effect. If there is large variation, it will be
hard to derive the underlying variogram from them.

So, if you modelled the variogram from the data assuming constant
measurement error, and then you add it (using `V' field) later on,
you're accounting for it twice.
--
Edzer