Philipp,
> we have performed a VBM analysis giving as a cluster of about 1400
> voxels at p uncorr 0.001, with a cluster corrected p-value of <
> 0.001.
>
> (The design was a two-group comparison (quite imbalanced: 40 : 140),
> with 3 covariates and a constant; preprocessing incl 12 mm smoothing
> and estimation in SPM5).
>
> We also used SnPM (version 3, in SPM2) to gain permutation based p-values,
> but now are stuck - statistically and technically.
>
> Questions:
>
> 1. Reading that cluster corrected p-values are not valid for VBM
> data due to non-stationary smoothness, what is the alternative? Are
> we going into the right direction by gaining NP p-values? Is the
> non-stationary smoothness problem somewhat 'compensated' by the
> large N?
This is not clearly documented, but SnPM's FWE-corrected P-values
*are* valid under non-stationarity. FWE P-values are based on the max
distribution, and the max implicitly accounts for differences in
smoothness. It does mean that areas with less smoothness will have
reduced sensitivity relative to a non-stationary test (see Moorhead et
al, NI 28(3)-544-552), but the FWE P-values are valid.
The *uncorrected* cluster size P-values *do* make an assumption of
homogeneous smoothness.
> 2. SnPM (after 5000 permutations) showed a more or less identical
> cluster after thresholding the uncorr. p-values at 0.001. We tried
> to collect spruathreshold clusters, but when we try to view the
> results an error appears 'SnPM_ST.mat corrupt', i. e. this file
> cannot be read in (3.5 GB).
You need to specify the cluster-defining threshold *first*, before the
computataion stage. This SnPM3 feature will avoid the creation of the
SnPM_ST.mat file.
> 3. Do you have any alternative suggestions how we can gain a p-value
> for the whole cluster?
Stick to the FWE-P-values... is the signal robust enough?
> 4. Looking at the ResMS image of SPM, we see that the cluster lies
> in an area of relatively high residuals (compared with other
> areas). May be this roots in more residual gyral variation in this
> area after spatial normalisation. Is this issue 'solved' by using a
> permutation method?
Residuals aren't so much the issue as local *smoothness*. You can
look at an image of FWHM by transforming the RPV image (with ImCalc)
with the following expression
FWHM = (RPV)^(-1/3)
The concern is: Do the clusters tend to fall in the regions of
relatively greater or less smoothness. If they fall in relatively
rough regions, a stationary method is probably underestimating their
significance, if they fall in relaticely smooth regions, a
stationary method would overestimate their significance.
Does this help?
-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
|