I'll give it a try...
If you see this document by Jeanette Mumford (http://mumford.bol.ucla.edu/mult_test_2009.pdf)
and go to slides 11-13, this should help. What it means is that, if the T value for p<0.01 is, say, 3.5, then the cluster-level p-values give the probability of getting a cluster at least as large as k voxels, given that you've already thresholded the imaged at T > 3.5. If the p-value you care about is .05, and one of the cluster-level p-values is below that, then you reject the null hypothesis of no activation in the cluster; thus you can't say the whole cluster is "active", just that somewhere in there lies an active voxel.
Is this correct, experts?
________________________________________
From: SPM (Statistical Parametric Mapping) [[log in to unmask]] On Behalf Of Graham Williams [[log in to unmask]]
Sent: Wednesday, June 23, 2010 7:32 PM
To: [log in to unmask]
Subject: [SPM] Thresholds for visualization of clusters vs. voxels
Using SPM2, I set the p value for threshold in the Results t-test contrasts to 0.01. To my understanding, this displays all voxels that are significant at the p < 0.01 level. However, in the Volume statistics multiple clusters are reported which have p values much higher than 0.01. Why is it then that if I can see the voxels (which are presumably <0.01) in these clusters, the cluster itself has a much higher p value (uncorrected). I am wondering whether this is due to the far more extensive variance in a cluster or whether I do not fully understand the threshold setting.
Thanks for any help you can give, Graham
|