Hi everyone,
I’ve run into an analysis question that I don’t
understand. I am looking over books and the email list but I’m wondering
if anyone has any ideas to at least get me started in the right direction.
I’ve done two analyses on one individual. The data
were acquired with 3.75 x 3.75 x 6 mm voxels. In the first analysis, I
normalized to the standard brain and resampled to 3 x 3 x 3 mm voxels. In the
second analysis, I skipped the normalization and maintained the original voxel
size. In both analyses, I applied a smoothing kernel of 7.5 x 7.5 x 7.5 mm FWHM.
In fact all of the other preprocessing steps and options in the statistical
analysis were exactly the same. The only difference is which data I input into
the model (normalized or unnormalized).
The SPM results indicate a smoothness of 31.0 x 35.5 x 27.2
mm FWHM for the unnormalized analysis, and 14.6 x 14.7 x 16.4 mm FWHM for the
normalized analysis (does this mean my unnnormalized images are smoother? This
is confusing to me, it seems to me that the upsampling done in the normalization
step would have a smoothing effect as voxels are interpolated).
The major concern is that I lose a lot of power in the
unnormalized analysis. The two analyses look about the same when I set the
threshold to p <0.05 FWE corrected in the normalized analysis and p <
0.05 uncorrected in the unnormalized analysis.
I had a look at my ResMS.img’s and they are of much greater
intensity in the unnormalized analysis, which I think is contributing to my
loss of power. The problem is I don’t understand how normalization would
have this effect, unless there is some assumption inherent to SPM that
normalized data and/or isotropic voxels are being inputted. Based on reading
the email list this doesn’t seem to be the case though.
Any help/insight/hints would be greatly appreciated!
Thanks very much,
Erin