Dear Users,
I apologize to resubmit this question to your brains, but I need to
understand some of my results to be sure that I did not make errors setting
my analysis.
So, even (and particularly) if the answer is "You are totally wrong", please
let me know !
>From previous analysis (semi-variograms using package gstat), I found
spatial autocorrelation in my dataset.
The best fitted model to this spatial correlation structure is the Gaussian
model (Spherical, Exponential, Linear tested and comparison done by Sum of
Square errors).
So I used corGaus function to define this spatial autocorrelation in my lme
model using the option "correlation".
Example: lme( Y ~ X1 + X2 + ... + X3, random = ~ 1 | RandomFactor,
correlation = corGaus(form = ~ X + Y | RandomFactor )
The Variogram function (package nlme) used on a lme object calculates the
semi-variogram for the within-group residuals and add the semi-variogram of
the corSpatial element (corGaus in my case) included in my model ... so far
no problem.
I was surprised, however, to see on the plot of the semi-variogram issued
from the Variogram function, (see figure at http://imm.io/3OLe) the low
range value (~1600 meters) used in the corGaus structure included in the lme
object.
When I fitted the same corGaus structure manually or using the fit.variogram
function (package gstat) on the data of each group defined in lme, it gaves
me ranges between 2050 and 2700 meters (mean 2350 meters).
Can anyone explain me those differences ?
Note: As I mentioned in a previous message (
http://markmail.org/message/gjgag4ohjopevgax), I tried to define a different
range in the corGaus function directly in the lme function, but it is not
taken into account.
Arnaud
[[alternative HTML version deleted]]