Hi Ling,
Looking at the log files for a Level 2 analysis, smoothest gets run
again by FEAT at Lev2.
e.g., here is the command for a Lev2 analysis involving 4 inputs, and a
two column design matrix (thus 2 d.f'.s):
/usr/local/fsl/bin/smoothest -d 2 -m mask -r stats/res4d >
stats/smoothness
So, it would appear that the smoothness values used in a Level 2
analysis within FEAT are those from the actual residuals of the Level 2
analysis, not the average of the values from the inputs at Level 1.
As to whether averaging the level 1 values can be (and under what
conditions) a reasonable approximation to the actual smoothness at level
2, I'll defer to someone more knowledgeable about Gaussian Random
Fields.
Best,
-MH
On Thu, 2011-02-24 at 00:57 +0000, Josef Ling wrote:
> Hello-
>
> We are trying to use easythresh to do family-wise error correction based for level two analyses. The stat maps were generated in another program for an experiment that had 27 subjects. We would appreciate some feedback to see if the following approach is valid.
>
> 1) First we obtain smoothness estimates for each subject’s residual time-series following first level analyses (time-series = 149 images)
>
> smoothest –d 148 –r sub1.resid.nii.gz –m ave_mask -V
> …..
> …..
> …..
> smoothest –d 148 –r sub27.resid.nii.gz –m ave_mask –V
>
>
> 2) We then average the temporally corrected DLH and RESELS values from all subjects to obtain a single DLH ($ave_DLH) and RESEL ($ave_RESEL) for the experiment. Is this similar to what is done in FEAT for level two analyses?
>
> 3) Finally, we use easythresh to correct for false positives, after first hard-coding our derived DLH and RESEL values from step 2 into the easythresh script.
|