Klaus and Karl,
I, too, have now tried the multi-subject, covariates only approach, but
this approach does not produce results that match a simple correlation of
covariate difference scores with image difference scores. Rather, the results
using multi-subject, covariates only is very close to using conditions and
covariates, with no specific fits.
As I have said before, the only way I can generate results in SPM96b
that match a simple correlation of difference scores with difference images is
to use conditions and covariates with a condition specific fit for the covariate
and a 0 0 -1 1 contrast.
Just to be clear, the covariate is constructed by taking the difference
score of each subject (task - baseline condition), calculating the mean of the
difference scores across all subjects, subtracting the subject score from the
mean of difference scores, dividing the mean-centered score by 2 for each
subject, and finally multiplying the result by -1 for the control (baseline)
condition and +1 for the active (task) condition.
This leads me to conclude that when if the covariate is colinear with
the conditions, as with a task performance score covariate, then it is necessary
to explicitly model the condition by covariate interaction to get the equivalant
of correlation of difference scores with a difference image.
I think much of the confusion on the subject of analyzing covariates of
interest, therefore, is due to a failure to explicitly state whether the
covariate is colinear with the condition. If the covariate is not colinear then
all of the approaches should give similar answers. If the covariate is
colinear, then each approach will give a different answer.
Karl, is this a correct conclusion ?
Also, in contast to Klaus, I find that inclusion or exclusion of a
covariate in the model has no influence on the contrasts that only look a
condition effects.
sg
Dear Karl,
you responded to:
> In order to examine the effect of the covariate on brain
> perfusion/metabolism within subjects in SPM (which I believe is what
> correlating difference scans with difference psychometric measures
> does) you would have to model subject (to remove intersubject
> variance), not condition, but covariate, which would in the absence of
> a condition effect acount for the variance within subjects.
>
> If this is true, how do you configure the model within SPM to model
> only subject and covariate effects (remember there are two scans and
> two covariate measures per subject) without modelling condition?
with:
Simply use multi-subject, covariates only.
------
I have done this and it works. (SPM 96 and SPM "97" Windows).
Unfortunately my dataset gives identical results, when I use 2
conditions or 2 replications (1 condition). Which leads me to the
question: in which order are the various effects entered into the
model. Or in other words, are there effects that are estimated only
after other effects have been modelled? For example, confounds
first, then covariates of interest, then condition and subject effects -
or all simultaneously? By playing with the data, I found for example
that SPM estimates of condition effects vary with the covariate of
interest entered into the model, i.e. they seem to be computed after
the covariate of interest effect has been taken into account (or
simultaneously). - Condition effects are also identical whether the
covariate is entered as a confounder or a covariate of interest.
Many thanks for you help and forbearance
Klaus
Professor KP Ebmeier
Department of Psychiatry
University of Edinburgh
Edinburgh EH10 5HF
United Kingdom
Tel/Fax: *44-131-5376505
Website: www.pst.ed.ac.uk
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|