S-News Readers,
Does anyone have experience in using lme() and corSpher() or corExp() to
fit linear models with a spatial covariance structure using REML? I
have been using S+SpatialStats and nls() to fit variogram models using
weighted least squares. I usually fit a linear model first using lm()
or glm() and then use the residuals to fit a spherical variogram model
using nls(). However, I'd like to try fitting the linear model and
covariance structure simultaneously using REML. The lme() and
corSpher() functions seem like the way to go, but they're new to me, and
I have questions. First of all, how do I get the standard errors for
the spherical model parameter estimates: range, nugget, and sill? The
standard errors are the main reason why I wanted to try using REML in
the first place. Secondly, when I use the corSpher() function along
with the "nugget=T" argument, I don't get an estimate for the sill. The
sill is one of the parameters for the spherical variogram model, so why
don't I get the estimate for it? Finally, I'd like to use a likelihood
ratio test (LRT) to test whether the spatial autocorrelation in the
residuals from the linear model is significant. It would seem like I
could use update() along with lme() to fit a model with and without the
spherical correlation structure and then use anova() to do the LRT.
However, the lme documentation warns against using REML and recommends
ML instead. So do I need to fit the models using ML in order to use the
likelihood ration test and not REML?
Thanks everyone for your time and patience.
-Paul
Paul Schwarz
Department of Natural Resources
Cornell University
Fernow Hall
Ithaca, New York 14853-3001
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|