Skip to content
Prev 3429 / 29559 Next

question about regression kriging

Edzer,

I do not want to get into too much discussion (this is definitively not a place to discuss basic
mathematical/statistical theories). My experience is - things are complicated when trying to run
regression-kriging with 0/1 variables. 

In practice, I guess you can fit GLM model, then estimate the variogram for residuals. Imagine that
we have a point map showing binary observations (0/1) and a set of raster maps PC1-PC5 (pc.comps).
Then, we can make predictions using glms and your package gstat:

# regression (GLM) modelling:
# variogram modelling (original scale):
# final predictions (A) regression-kriging:
model=vt.rbvgm)
Now you have the predictions that can exceed 0-1 range, and no estimate of the UK variance (this is
the problem David has, I think). What we can do instead is to run UK in gstat on logits first, then
back-transform to original 0-1 scale: 

# Estimate the binary variable at logit-scale (as an average of the output of GLM and original 0/1
value). This is not possible if values are equal to 0/1, so we have to 'smooth' them:
# Now the values can be back-transformed to logit scale
# now fit variogram for residuals (logits):
# final predictions (B) UK in gstat:
model=vt.logitvgm)
# Back-transform (-20 to 20 values) to original scale:
Obviously, predictions with (B) are more attractive than predictions with (A), because the values
are within the 0-1 range and we also have the UK variance (but this is hard to interpret). Then, one
can standardize the prediction variance using the global variance so that you can do some
interpretation:
Of course, (B) is only a short-cut solution. One really needs to develop a sound methodology to
solve such problems (we at least offer a detailed discussion in
http://dx.doi.org/10.1016/j.geoderma.2007.04.022).


PS: How do you back-transform the GLM prediction variance to original scale (I was not aware of
this, apologies)? A reference would do.

Tom


-----Original Message-----
From: Edzer Pebesma [mailto:edzer.pebesma at uni-muenster.de] 
Sent: woensdag 9 april 2008 10:49
To: Tomislav Hengl
Cc: 'David Maxwell (Cefas)'; r-sig-geo at stat.math.ethz.ch
Subject: Re: [R-sig-Geo] question about regression kriging

Tom,

I'm afraid things are harder than you sketch. In glm's, the parameter 
estimation is done using iteratively reweighted least squares, where the 
weights depend on a variance function that links the variance of 
observations to the mean. So, observations (residuals) are assumed to be 
unstationary, in principle, and because of the mean-dependency this 
changes over the iterations. The equations and references you mention 
afaik all assume a known, and fixed variogram, and one-step solutions, 
no iteration.

Also, you falsly accuse me of claiming one cannot back-transform 
prediction variances. I did not claim this (I have seen suggestions on 
how to do this), I just asked how David would do this.
--
Edzer
Tomislav Hengl wrote:
wrong
Cressie
Earth
(logits).
Of
geoRglm
there