A simple answer would be that the cluster-level correction accounts for
all the clusters in your statistic image. Typically, the number of
voxels in a statistic image can be 10,000+, and a voxel-level multiple
comparison correction needs to account for 10,000+ multiple comparisons.
On the other hand, there are at most a few hundred clusters occurring in
a statistic image, so a cluster-level multiple comparison correction
only needs to account for a few hundred multiple comparisons.
I hope this makes sense.
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Jeff Browndyke
Sent: Thursday, September 25, 2008 3:35 PM
To: [log in to unmask]
Subject: Re: [SPM] Cluster corrected p-values. By what means?
Thanks for responding. We typically set a more stringent threshold of
20k for 3.75mm voxel size to buttress against false-positives when FDR
or FWE don't hold.
I was wondering more specifically about the routine SPM5 results output
where clusters are assigned in the table a "corrected" and "uncorrected"
p-value regardless of the voxel-wise correction. These p-values are
noted to the left of the voxel-wise columns.
I was under the assumption that if FDR or FWE were not reached at the
voxel-level, you could at least report ONLY those clusters that reached
a good cluster-wise significance threshold increasing confidence in your
results. The "uncorrected" column is simple enough, but what's the
"corrected" column for cluster significance reflecting? (Size of
cluster relative to the sizes of the other clusters / total
clusters)...or something like that??
---- Michael T Rubens <[log in to unmask]> wrote:
> the extent threshold is not factored into the correction, i.e., it
> would be the same whether the cluster is 1 or 100. it just changes the
> output for visualization, and some groups have a standard threshold of
> 5 or 15 voxels. If you want cluster extents to factor into correction,
> check out RFT (random field theory).
> On Thu, Sep 25, 2008 at 8:45 AM, Jeff Browndyke <[log in to unmask]>
> > Hopefully this question isn't too simplistic, but in the SPM output
FWE and FDR are explicitly stated for voxel-corrected p-values, but what
is the means of "corrected" for the SPM "non-isotropic adjusted
cluster-level" p-values? And, is the "set-level" p-value corrected
already, if so by what means?
> > Thanks for any education in this regard...the p corrected at the
cluster-level has been bothering me for a while, but now enough to seek
a definitive answer.
> > Regards to all,
> > Jeff Browndyke
> Research Associate
> Gazzaley Lab
> Department of Neurology
> University of California, San Francisco