[R-sig-Geo] spatial covariance in the glmm in brain data

[Michael Klein]

Dear List,

I am a beginner or R (just started last week) and I am not exactly a freak
in theoretical statistics, so please excuse me, if my questions are naive or
contain mistakes. This might also not be your standard problem since it
is about brain imaging data. However, it seems that I have to deal with a
problem that you seem to be very familiar with.

I try to model time frequency data with the glmm while either correcting
for spatial covariance or removing the spatial covariance somehow.
More precisely, I try to model the mean power change for time frequency
bins by the fixed factors TIME and FREQ and the random factor SUBJECT.
Since in time-frequency data, the values of time frequency bins are not
spatially independent, I would like to either remove that dependence or
correct for it.

I must have read a dozen papers and pages and posts and came up with
something that might be going in the right direction. However, I need to 
be sure
that this is doing what I want it to do:

My data (tfbins, value: mean_change) is basically the result of an 
average for
every subject for each time freq bin and then the percentage signal change
between the two conditions. So for every subject I have one time freq 
matrix
(everything transformed to a data frame).

To model the spatial covariance, I understood (maybe incorrectly) that I 
can use
a corExp term. So here is what I came up with. Instead of x and y spatial
coordinates, I just use time and freq:

model_spatial_2 <- glmmPQL(mean_change~time*freq, random = ~1|sub, 
correlation=corExp(form=~time+freq), data = tfbins, family = "gaussian")

Any thoughts on that would be very much appreciated.

Best regards,
Michael