Hi Darren, (and SPMers),
Let me cross-post this response to the SPM list. I think it might be
helpful for others too.
> First would non-stationary RFT methods apply primarily to cluster-size
> thresholds?
Yes, that's correct.
> From your paper in 2004 non-stationary RFT methods perform well for
smooth
> images under high degrees of freedom. For fmri I hope this would
> correspond
> to first level analyses where smoothing was double the voxel size
> (although
> I see in the paper that smooth mean at least 4 voxel FWHM).
We found that, in general, RFT (stationary or non-stationary) performs
well with df~30 or larger. So this applies to the first level for sure.
If you have enough subjects (30+), probably RFT performs fine in the
second level too.
> What about at the second level? My understanding is that
non-stationary
> RFT
> tests would lose sensitivity in this instance because of the low DF.
> However, non-stationary RFT tests have better sensitivity that
stationary
> tests in rough regions. This leaves me confused as to the "best'
choice if
> indeed there is a best choice.
My guess is that, for the second level fMRI analysis, you probably would
do well with the regular stationary RFT in SPM than the non-stationary
RFT. The non-stationary RFT introduces additional uncertainty to the
statistical test because the smoothness has to be estimated at each
voxel; smaller the df is, less precise the smoothness estimate becomes.
My guess would be that reduced sensitivity due to smoothness estimation
would probably overwhelm the benefit of increased sensitivity in "rough"
areas.
Now, the story is different in VBM-type analyses though. I have often
observed clear and systematic structural variability in local smoothness
in RPV (resel-per-voxel) images from VBM analyses. So, although
non-stationary RFT reduces the sensitivity, it would be more appropriate
in VBM analyses in order to minimize sensitivity biases in smooth areas.
> Would one choose stationary RFT if primarily looking at voxel-level
stats
> and non-stationary RFT when looking at cluster level stats?
Sure. In fact, our NS toolbox calculates the voxel-level stats exactly
the same way as the regular RFT. The reason behind this is difficult to
explain, but I had a discussion with Ged Ridgway in the past about this
topic. See
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind06&L=SPM&D=0&I=-3&m=24842
&P=457146
> Do you have any general guidelines about these choices?
Well, in a nutshell, I suggest doing the regular stationary RFT for
second level fMRI analyses, unless you have a strong reason to believe
that the smoothness is dramatically different in different areas of the
brain. You can see this by looking at the RPV image (RPV.img, a
byproduct of an SPM analysis). Perhaps an FWHM image is more
interpretable than an RPV image. You can transform RPV to FWHM quite
easily. See
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind06&L=SPM&P=R333724&I=-3
Hope this helps!
-Satoru
|