[log in to unmask]"
type="cite">
Hi Anderson,
I ran a couple more analyses, 8 in total:
(data demeaned yes/no) * (design.mat demeaned yes/no) * ( -D
in randomise yes/no)
The dataset is very small (8 subjects) and thus, all
permutations in randomise were executed.
I compared the resulting checksums of the tstat and p-value
maps:
As you can see from the figure:
- p-values were the same for all models with the -D option
- p-values were the same for all models without the -D option
- tstats were the same for all models with the unaltered data
file
- tstats were the same for all models with the manually
demeaned file
However, I would have expected for example OPTION 07 and
OPTION 08 to be similar because regardless of the -D option,
both use demeaned data as well as a demeaned matrix.
Why is this not the case?
Thanks,
- Vincent
Op 11 mrt. 2014, om 18:26 heeft Anderson M. Winkler <
[log in to unmask]>
het volgende geschreven:
Hi Vincent,
Could you paste here the design.mat and the design.con files?
Another thing is which image are you using to compare the
results? The comparison must use the images of the statistic,
not of the p-values, because even with the same seed, there is
a possibility, depending on the design, that the shufflings
aren't exactly the same (except of course in the exhaustive
case), then the p-vals may vary a bit.
Thanks.
All the best,
Anderson
Am 11.03.14 21:22, schrieb Vincent:
I have VBM data: 1 group with 1
continues covariate.
I am interested in the association between the covariate and
local gray matter volume.
From what I understand from Jeanette Mumford's website (http://mumford.fmripower.org/mean_centering/),
I do not need to mean center the covariate, and according to
the FSL GLM wiki (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Average_with_Additional_Covariate)
I should use the -D option.
If I understand correctly the -D option demeans both the
design matrix as well as the data (FSL v 5.0 and later). To
get a better understanding about what -D is exactly doing, I
ran randomise two times:
1) On a regular 4D nifti image with smoothed gray matter
images in MNI space:
randomise \
-i 4D_GM.nii.gz \
-o vbm \
-d design.mat \
-t design.con \
-m ../mask.nii.gz \
-n 500 \
-T \
-1 \
-D
2) On the same 4D nifti file that I manually demeaned:
fslmaths 4D_GM.nii.gz -Tmean mean.nii.gz
fslsplit 4D_GM.nii.gz
for i in `ls vol*`; do fslmaths ${i} -sub mean.nii.gz
new_${i}; done
fslmerge -t 4D_GM_demeaned.nii.gz new*
with a manually demeaned covariate in the design matrix
other files were kept constant (design.con, mask.nii.gz)
randomise \
-i 4D_GM_demeaned.nii.gz \
-o vbm \
-d design.mat \
-t design.con \
-m ../mask.nii.gz \
-n 500 \
-T \
-1
The results from these two methods were not the same.
When I subsequently added the -D flag to the second run with
the manually demeanded data, the results were the same as
with those of the 1st model.
Could anybody please explain the discrepancy between the
results of 1) and 2)?
- Vincent