Lena,
> I have noticed that when I compared concentration (i.e. unmodulated)
> images between two groups of subjects, the values at the bottom of
> the output for smootheness that is estimated by SPM99 was:
>
> FWHM=11.7 13.7 12.6 mm (which is close to the 12mm kernel I used to smooth
> the images).
> voxel size = 5.9 6.8 6.3. ( The voxel size for my concentration images is
> 2x2x2.)
>
> When I re-run the same analysis using volume (i.e. modulated) images
> derived by applying the optimised protocol (as described by Good et
> al), the estimated smootheness was :
>
> FWHM = 20.5 20.9 19.9 mm. (I smoothed the volume images with 12 mm kernel)
> voxel size = 20.5 20.9 19.9 mm ( The voxel size for the volume images is
> 1x1x1).
> 1. Why would the estimated FWHM and voxel size be the same for the
> volume and both be very different from the 'real' values; but near
> the 'real' values for the concentration?
The smoothness is estimated on the standardized residual images. The
only way for estimated smoothness to go up is for the inherent noise
to become smoother *or* for there to be increased structure in the
residual images.
I've never looked at modulated images, but I won't think they have
dramatically different smoothness than unmodulated (John?). If
smoothness is similar, that leaves the second explanation, that there
is "more stuff", greater spatial structure in the residual images.
The standardized residual images are essentially each subjects data
after subtracting off the intragroup mean image (and dividing by
standard deviation), so this suggests that subject's anatomy exhibits
greater differences with modulated than with unmodulated images.
This is my take. I'd be interested in what John et al. think.
One thing you can do is look at the two studies' RPV images, these are
the images of Resels Per Voxel, or roughness. Also, look at the
character of the residual variance, ResMS. Is there different or
weird structure in the RPV image of one study but not in another?
(Actually, both RPV and ResMS are poorly scaled for visualization. I
always use ImCalc to create
ResRMS = sqrt(ResMS)
rRPV = sqrt(RPV)
images and look at these instead. And to quantitatively compare across
studies, you may want to look at the coefficient of variance from each
study,
CofVar = sqrt(ResMS)./beta_000x
where beta_000x is the beta image corresponding to the grand mean.)
> 2. How does this estimated voxel size of ~ 20 mm voxel size in all
> directions reflect on the statistics for the between group volume
> comparison, i.e. does this make the stats more conservative (or more
> liberal)?
All things equal, increase smoothness increases corrected
significance. So this increase in FWHM smoothness will lower
corrected thresholds and corrected P-values.
> 3. Related to that, if I choose smaller smoothing kernel for 1x1x1 voxel
> size data, would that result in more sensitive (or more conservative)
> statistics?
Likewise, all things equal, decreasing smoothness decreases corrected
significance.
Hope this helps.
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|