> I am confused as to the meaning/what SPM is doing when you do/do not use FWE. For example, for the attached output, I used FWE p = 0.05 with threshold = 0. The cluster that survived the FWE correction has a pFWE = 0.01, p_uncorrected = 0.198- what does this mean?
I think perhaps the confusion is that SPM gives you the option to
correct for multiple comparisons at different levels—in this case, at
the voxel level or at the cluster level. Both are perfectly valid,
but permit inference at different spatial scales. A voxel-level
correction lets you say that voxel X is unlikely to occur by chance,
and therefore informative regarding your hypothesis. A cluster-level
correction says that *cluster* Y is unlikely to occur by chance (and
therefore that 1 or more voxels within that cluster are unlikely to
occur by chance—but not which voxels in that cluster).
The p value that you set when running the results command is the
voxelwise threshold. At the voxel level, this is your p value
(corrected or not). From a cluster perspective, this value is the
threshold that defines what the clusters are, and is thus often
referred to as the "cluster-defining" or "feature defining" threshold.
So, when you set FWE p = .05, you have voxels that are significant (at
the voxel level) p < .05. You have also defined a cluster in the
attached output, which also produces some cluster-level statistics.
Generally if you use a stringent voxelwise correction (as you have
done) you don't additionally need to look at the cluster-level
correction, because by definition every voxel in your cluster is
unlikely to occur by chance. (In this particular case it's odd that
the FWE p value for the cluster is lower than the uncorrected cluster
value; I'm not sure why that is.)
> For other analyses, I've first tried FWE p = 0.05, did not obtain any results, then did not choose FWE and set p = 0.001. The results showed clusters with pFWE < 0.05; why didn't they survive the initial FWE p = 0.05 correction? Does it mean something different when FWE is not used, but there are clusters with pFWE < 0.05?
In this case, you've set a cluster-defining threshold of p < .001
(uncorrected), which gives you some clusters. Based on their extent,
these clusters may or may not be bigger than expected by chance. In
this case, you have some clusters that were indeed larger than would
be expected by chance, giving you a cluster-level p values of < .05
(even though there were not any voxels that survived a voxel-level FWE
correction). With this result, you would be entitled to make
inferences at the level of clusters, but not of individual voxels
within that cluster.
Hope this helps!
Dr. Jonathan Peelle
Department of Neurology
University of Pennsylvania
3 West Gates
3400 Spruce Street
Philadelphia, PA 19104