Date: Tue, 29 Jan 2008 20:17:42 +0100
From: "G. Allegri" <giohappy at gmail.com>
Subject: Re: [R-sig-Geo] Regression kriging
To: "Edzer Pebesma" <edzer.pebesma at uni-muenster.de>
Cc: r-sig-geo at stat.math.ethz.ch, Jose Funes <jefunes at gmail.com>
Message-ID:
<e12429640801291117j4f31de22iae9d03bbeb6d4218 at mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1
Dear Edzer,
I've "medidated" on the answer you gave to Jose. Two considerations have raise:
1 - when you say that the approach of GLM is a way to consider
spatial dependence. I'm not sure about this. GLM are a way to account
for link functions between the dependent variables and covariates (ex.
Poisson family for count datas), but they don't take account,
implicitly, of sptial correlation. Am I wrong?
Rather (generalized) mixed models are a counterpart to geostatical methods are.
2 - A task of my research is to find the "best" relations between a
set of covariates, to make a simple multicriteria analysis,
overlapping different map layers thorugh map algebra. In this case,
the common geostatistical methods don't help me much. I'm considering
to use multivariate regression, but keeping in count of spatial
correlation. What's the best approach? I've thought to Mixed Models,
but another way could be using GLS estimation, based on the residauls
covariance. What's your suggestion?
Giovanni
PS I think it could be an answer to Jose too...
2008/1/27, Edzer Pebesma <edzer.pebesma at uni-muenster.de>: