Dear Joe,
>>we appear to have run into a possible bug in SPM2's cluster-level
>>correction routines. As you can see in the attached output files, if
>>we reduce the height threshold below about .001 the corrected
>>cluster-level p-values actually become more significant than the
>>uncorrected p-values. The first file is at .001 and is okay. The
>>second file is at .0001 and is not okay. We've seen even single voxel
>>clusters declared significant by the corrected p-values even though the
>>uncorrected p-values are near unity. This is happening pretty
>>consistently in our analyses. This can't be right, no? I can't think
>>of how an MCP correction could be anything other than more conservative
>>than an uncorrected test. Or are we doing something wrong?
This is a very good point. The explanation is important but subtle.
The uncorrected p-values based on spatial extent are conditional on
the cluster existing (p). The corrected P values are not, they
are simply the probability of getting one or more clusters of that
size or greater in the search volume (P). In fact P = Em*p, where
Em is the expected number of clusters. Given that Em is usually
small (0.05) you can see that p > P.
The reason for this slightly odd reporting of extent-based p values
is that the uncorrected p value can be used for randomly selected
clusters (which, by definition, exist). This is described more fully
in:
Friston KJ. Testing for anatomically specified regional effects
Hum. Brain Mapp. 1997;5(2):133-6
I hope this helps - Karl
|