Dear Shary and SPMers,
At 09:51 PM 8/2/2005 +0100, Shahryar Rafi-Tari wrote:
>I am wondering how the expected values shown on the SPM2 output relate to
>the SPM results. How the following are defined:
>
>1)- Expected voxels per cluster
>2)- Expected number of clusters
>3)- Expected false discovery rate
These are expected values, (i.e., means) based on random field theory.
These parameters are used to determine FWE-corrected voxel p-values (2) and
cluster p-values (1 & 2), as well as FDR-corrected voxel p-values (3, I
think).
>For instance, in an SPM2 output I got the following after setting voxel
>threshold to 0.001 uncorrected and extent threshold to 100 voxels:
>
>1)- Expected voxels per cluster =96.863
>2)- Expected number of clusters= 0.77
>3)- Expected false discovery rate <=0.42
>
>I know from experience that if I set voxel threshold to 100, the sizes of
>the displayed clusters will be at least 100 voxels. But, why I am getting
>expected number of clusters =0.77? In this output I have one cluster of
>the size 157 voxels and corrected p=0.39 and uncorrected p=0.18.
You are right, if you set the cluster size threshold to 100, then only
clusters with 100 voxels or larger are displayed. However, in the
calculation of the expected voxels per cluster (i.e., expected cluster
size) is separate from the cluster size threshold you enter. In fact, the
cluster size threshold you enter does not influence your cluster size
p-values whatsoever, but it is used in the calculation of the set-level
p-value only.
The expected number of clusters is 0.77 doesn't mean no cluster should
occur. This is basically the mean number of clusters. If the images
contained only Gaussian noises, this would be the number of clusters one
would expect to occur.
>The expected discovery rate makes sanse as the output shows it. But, how
>the expected voxels per cluster and expected number of clusters relate to
>the height threshold and extent threshold.
Well, in a nutshell, these parameters are used to calculate the probability
density function for cluster size based on random field theory. The cluster
size distribution is not a simple distribution, but if you are interested,
you can see how it's calculated in
- Friston et al. Human Brain Mapping 1: 210-220. (1994)
- Worsley et al., Human Brain Mapping 4: 58-73 (1996)
- Hayasaka & Nichols. NeuroImage 20: 2343-2356. (2003)
-Satoru
Satoru Hayasaka ==============================================
Post-Doctoral Fellow, MR Unit, UCSF / VA Medical Center
Email: shayasak_at_itsa_dot_ucsf_dot_edu Phone:(415) 221-4810 x4237
Homepage: http://www.umich.edu/~hayasaka
==============================================================
|