 |
 |
SnPM finds the distribution of the maximum, whether it be a cluster size or voxel intensity, and calculates FWE corrected p-values based on this distribution. But that does not mean the corrected p-values only apply to the maximum. Finding the distribution of the maximum is another way of controlling multiple comparisons, since P(Kmax>=k) is statistically equivalent to the probability of finding one or more clusters with k voxels. This rationale is explained in details in Andrew Holmes's original SnPM paper in JCBFM in 1996, and by other papers since then (Bullmore et al, IEEE-TMI (1999); Nichols & Holmes, HBM (2002); Hayasaka & Nichols, NeuroImage (2003); etc). -Satoru -----Original Message----- From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Rajeev Raizada Sent: Tuesday, August 26, 2008 4:43 PM To: [log in to unmask] Subject: [SPM] Isn't max-cluster-size meant to be a property of whole brain, not of a cluster? [Re: Q: Show P(Kmax>=k) in SnPM results figure?] On Tue, 26 Aug 2008 11:57:46 -0400, Satoru Hayasaka <[log in to unmask]> wrote: >> I am using SnPM to perform multiple-comparisons correction, >> by calculating the p-val of getting the observed >> maximum suprathreshold cluster size. >> I would like to display this corrected p-val in the output figure. > >The SnPM online manual is somewhat out of date. In your SnPM results, >you should look at the column under p_FWE-corr for corrected p-values >for clusters. Many thanks for the mail. If I understand SnPM's maximum suprathreshold cluster-size p-value correction correctly, then this observed maximum size is meant to be a property of the whole brain, not of any individual cluster. i.e. "Given a whole brain of voxels with permuted labels, what is the prob that I would end up with any cluster anywhere in the whole brain whose size Kmax is greater than the value k?" The very fact that this gives us just one single p-value, for the whole brain, is, as far as I understand it, the reason why this max-cluster-size trick can be used for performing multiple comparisons correction. There is only one comparison made, with one p-value. The p_FWE-corr values listed in the results table that SnPM prints out are p-values for each individual cluster, as opposed to being the one single p-value, P(Kmax>=k), which tells me what the chance is that a brain full of permuted voxels would have given me a cluster as big as the biggest one that is actually observed. Whatever the p_FWE-corr values given for each cluster might be, they are not, as far as I understand it, the volume-wide P(Kmax>=k) that I would like to display. I might very well have misunderstood this, in which case please correct me! Many thanks, Raj