[R-sig-Geo] geoR Beta Estimates Not Equal to lm Coefficients: Why?

[Joshua Palmer]

Thank you very much Paulo and Edzer; between your responses and a little
brushing up on my geostatistics it now makes perfect sense that geoR uses a
GLS approach to removing the trend in likfit or krige.conv.  Otherwise, by
relying on OLS, I am not accounting for the spatial correlations in my
dataset (as you have suggested Paulo).  As Edzer has explained, if I were to
rely on OLS I am assuming there is spatial independence between my
measurements; therefore, my semivariogram model would be a pure nugget model
with zero range and zero partial sill.  My data clearly shows this is not
the case, which is why I cannot match my coefficients obtained from OLS with
those obtained by GLS.

Best Regards,
Joshua Palmer
Meteorologist
Atlanta, GA, USA

-----Original Message-----
From: Paulo Justiniano Ribeiro Jr [mailto:paulojus at c3sl.ufpr.br] 
Sent: Monday, September 03, 2007 12:45 PM
To: Joshua Palmer
Cc: r-sig-geo at stat.math.ethz.ch
Subject: Re: [R-sig-Geo] geoR Beta Estimates Not Equal to lm Coefficients:
Why?

Josha

If I've got oyur question correctly the diference is because geoR is
estimating both, regression coeficientes and covariance parameters and
therefore the estimates ar not the same as lm() which uses ordinary least
squares.
In other words the regression parameters on geoR are estimated considering
the spatial correlation between the sample points as described by the
assumed model

Hope this helps
P.J.


Paulo Justiniano Ribeiro Jr
LEG (Laboratório de Estatística e Geoinformação) Universidade Federal do
Paraná Caixa Postal 19.081 CEP 81.531-990 Curitiba, PR  -  Brasil
Tel: (+55) 41 3361 3573
Fax: (+55) 41 3361 3141
e-mail: paulojus AT  ufpr  br
http://www.leg.ufpr.br/~paulojus

-----Original Message-----
From: Edzer J. Pebesma [mailto:e.pebesma at geo.uu.nl] 
Sent: Monday, September 03, 2007 2:47 AM
To: Joshua Palmer
Cc: r-sig-geo at stat.math.ethz.ch
Subject: Re: [R-sig-Geo] geoR Beta Estimates Not Equal to lm Coefficients:
Why?

Joshua,

My guess is that to estimate beta geoR uses generalized least squares and lm
and SAS ordinary least squares. If you use a pure nugget model, or some
other model with a range parameter sufficiently close to zero, i.e. model
the observations as independent, the estimates should be the same.
--
Edzer

On Mon, 3 Sep 2007, Joshua Palmer wrote:

> Hello everyone,
>
> I am using the excellent geoR package to perform Kriging with External 
> Drift.  As to be expected, for any given set of covariates, trend 
> coefficients I obtain performing least-squares regression in SAS equal 
> those coefficients obtained by using R's lm function.  However, those 
> trend coefficients do NOT match the estimates for the beta values when 
> I run krige.conv (or likfit or variofit) in geoR.  The coefficients 
> are often similar, but they are never nearly or exactly the same.  I 
> was under the assumption that geoR utilizes least-squares regression 
> (on the trend.d
> matrix) a la R's lm function to derive beta estimates.  This appears 
> to be incorrect.
>
> Would anyone be able to explain to me why I am noting this difference 
> or should I be noting a difference?  I have researched R mailing lists 
> and geoR documentation but have been unable to find an answer.  Please 
> let me know if additional specificity is needed.  This dilemma is 
> universal across different sets of covariates and covariance models or
parameters.
>
> Any assistance is greatly appreciated and I thank you for your time.
> Joshua Palmer
> Meteorologist
> Atlanta, GA, USA
>
> _______________________________________________
> R-sig-Geo mailing list
> R-sig-Geo at stat.math.ethz.ch
> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>