Dear Lazlo:
I think I've worked this out since then, and my oft repeated requests
to the group on this issue I think reflects my general
uncomfortableness at times with setting up statistical questions in
general.
I've also since moved the analysis to spm99 and reformulated it as
truly random effects (i.e. one difference image per subject) so the
analysis path described way below is not quite right. Here's what I
did. I wanted to understand if each subjects active-baseline
activation varied with respect to some covariate. I suppose you could
look at this as a subject by covariate interaction at the second
level.
So at the first level I put all the subjects into a large fixed
effects analysis modeling the effect of condition (active vs.
baseline for each subject). I then generated each subjects contrast
image (image of the parameter estimate) for the contrast of 1 -1
(assuming conditions entered as active then baseline). This produced
a set of con*.img images. For ease use I copied these to a new
analysis directory making sure to keep track of the order.
Then I ran a linear regression on the contrast images with reaction
time or whatever measure you want as my covariate. You don't have to
mean center this beforehand as spm99 gives you the choice to do that
in the analysis (though at some point the covariate needs to be mean
centered). That's it.
I have to say though quite remarkably having repeated this analysis
about 200 different ways that if the underlying stats are basically
the same then spm produces about the same result which is nice (i.e.
you can run it as linear regression or as single subject with a
covariate, etc.).
The only thing to remember about using contrast images is that they
are implicitly masked- i.e. they have NaN's (not a number) in place
of zeros outside the brain. If you use the basic stats section then
the scripts understand this. However, if you put the contrast images
into some other statistics then when it asks about masking use
absolute and -inf as the value, otherwise I think spm tries to
calculate a mask using the NaN's in the calculation. The other thing
is that assuming you've taken care of all global normalization at the
first level you should omit any global normalization or scaling at
the second level.
Hope this helps,
Darren
>Hi Darren,
>
>I'm a rookie with SPM and have the same problems, that you had in May
>99, concerning spm analysis with two groups/two conditions in a group
>and 1 covariate.
>I did not find any answer to your questions on the mailbase-system.
>Maybe you can help me ??? Were the contrasts you proposed for analysis
>right?
>
>(I copy the mail for you:)
>
>Dear SPM:
>
>Tried searching through the archives but couldn't find quite the answer
>I
>was looking for so here goes...
>
>fmri study
>2 similar study types.
>2 conditions per study (rest and active)
>1 subject reaction time score per study
>
>Did random effects analysis- to produce 2 images per study for each
>subject, used proportional scaling and hi pass filter.
>
>Entered into multi-study conditions and covariates design.
>
># studys- 2
># subjects in study 1 = 14
># conditions in study 1 = 2
>entered the scans
># subjects in study 2 = 12
># conditions in study 2 = 2
>entered the scans
>
>covariates = 1
>I took each subject's score (26 scores), mean corrected the lot and then
>entered it as
>
>-p1 p1 -p2 p2 ... (where pn is the subject's study specific and mean
>corrected score). I didn't divide by 2.
>
>No specific fits.
>No scaling
>No global norm.
>
>Contrasts
>The main effect of each study
>study 1 = -1 1 0 0 0
>study 2 = 0 0 -1 1 0
>
>The interaction with the covariate
>0 0 0 0 +1
>0 0 0 0 -1
>
>Is this correct- I really just want to see the effect of the covariate.
>
>Thanks,
>
>Darren
>Darren R. Gitelman, M.D.
>
>Mostly I am interested in the interaction with the covariates...
>
>Many thanks
>
>Dr. Laszlo Sturz
Darren R. Gitelman, M.D.
-------------------------------------------------------------------------
Cognitive Neurology and Alzheimer's Disease Center
E-mail: [log in to unmask] WWW: http://www.brain.nwu.edu
Voice: (312) 908-9023 Fax: (312) 908-8789
Northwestern Univ., 320 E. Superior St., Searle 11-470, Chicago, IL 60611
-------------------------------------------------------------------------
|