Dear Darren,
At 11:46 AM 10/20/98 -0500, you wrote:
| I want to make sure I am doing a random effects analysis correctly. I am
| using fMRI with block design task. There are 12 subjects and 2 tasks each
| with 2 conditions.
|
| The baseline condition of both tasks is basically the same, the active
| conditions differ- basically AB and A'B.
I presume you mean AB & A'C here...
Were the two tasks (AB & A'C) done in different sessions?
| Step 1) collapse images for each condition for each subject.
|
| Convolve HRF with boxcars: yes
| High pass filter: yes
| Global covariate: yes proportional scaling.
|
| Thus I end up with 4 images per subject.
|
| ------
|
| I then put this into a multi-study different conditions design and said
| there were 2 studies, 12 subjects per study and 2 conditions per study.
|
| Contrasts
|
| -1 1 0 0 For main effect of B vs. A
|
| 0 0 -1 1 For main effect of C vs. A'
|
| -1 1 1 -1 For interaction of (B - A) - (C - A)
|
| 1 -1 -1 1 For interaction of (A - B) - (A - C)
Ummm, I'm not too happy with this: You're mixing within subject and between
subject variance here.
| I tried also putting the data through multi-subject different conditions to
| look at the main effects but the activations were much more restricted than
| in the multi-study analysis and missed areas of activation that clearly
| showed up in individual subjects. Do I have to use this analysis for main
| effects?
The multi-subject different conditions analysis is more appropriate.
However, note that strictly speaking this is only correct for more than two
conditions under assumptions of variance sphericity.
Although areas may show up "clearly" in individual analyses, there may
still be considerable variability in their magnitude from subject to
subject, such that the between-subject variance incorporated in a random
effects analysis is quite large, and these seemingly coherent individual
subject effects don't result in a significant random effects analysis. MIPs
of individual subjects are especially misleading in this case, since they
show effect significance rather than effect magnitude (further since they
are individually scaled they don't show the variability of significance).
| If I had an anatomically restricted hypothesis about the interaction I
| assume it would be OK to threshold the map at p = .01?
...it would have to be *very* restricted: You'd have to pre-specify the
voxel, in which case you should only look at that voxel. I'd be lenient
enough to allow you to look at voxels within a FWHM of your target
location, but thresholding the entire image based on a point anatomical
hypothesis is asking for false positives elsewhere in the image - false
positives which will be too tempting not to try to interpret!
Hope this helps,
-andrew
+- Dr Andrew Holmes [log in to unmask]
| ___ __ __ Wellcome Department of Cognitive Neurology |
| ( _)( )( ) Functional Imaging Laboratory, Stats & |
| ) _) )( )(__ 12 Queen Square, Systems |
| (_) (__)(____) London. WC1N 3BG. England, UK |
+------------------------------------- http://www.fil.ion.ucl.ac.uk/ -+
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|