On Fri, Jan 27, 2012 at 4:35 AM, SUBSCRIBE SPM Anonymous
<[log in to unmask]> wrote:
Dear SPM users,
I have some questions about topological FDR.
1/ If defaults.stats.topoFDR is at 0, topological FDR can be applied?
No. It needs to be set to 1. Alternatively, you can remove defaults.stats.topo completely. I am working on a toolbox that does both topo and regular FDR corrections.
2/ I am not sure of how can we apply topological FDR on our results. I have seen in the forum an example: Use a T=3 to obtain the FWEc and the FDR p/c and then use the values when going through the results procedure again.
Is T=3 corresponded to p<0.001? How can we obtain a p=0.05? Is it correct to work with a p =0.05 for topological FDR?
T=3 corresponds to a p-value based on an equation. It does not correspond to p<.001. If you search for T to p conversion you can find a number of programs to do the conversion. These are all based on the cumulative distribution for t-distributions. This T-value is the voxel-wise p-value uncorrected for multiple comparisons. You do not want to use .05 for the voxel-wise level. There are a number of ways to achieve p=.05 at the corrected level.
Yes. it is correct.
In my results I have FWEp: 4.561, FDRp: Inf, FWEc:3641, FDRc: Inf! I don't really understand what I am supposed to do! Somebody can help me?
If I use the FWEc, is it correct? What is the difference between FDRp and FDRc?
All of these methods of correcting for multiple corrections are correct. FWEp means what is the voxel-wise t-threshold for an individual voxel to be considered significant after correcting for multiple comparisons using random field theory. FDRp means what is the voxel-wise t-threshold for an individual peak to be considered significant after correcting for multiple comparisons using topological FDR. In your case, no peak is significant using topological FDR. The cluster thresholds tell you how many voxels you need in a cluster for the cluster to be significant after correcting for multiple comparisons. I prefer cluster statistics myself as I don't believe the individual voxels could be significant. Applying an extent and voxel-wise correction is making the test more conservative, but is possible.
3/I don't understand the difference between p and q? It is in the cluster level and the peak level in whole brain: pFWE-corr? qFDR-corr?
p-value for p-value, which is standard in T-tests. p-value is the probability that something occurred by chance.
q-value for FDR because it is a different statistic. q-value is the number of positive tests you are willing to allow.
4/ What is the mean of FDRp: Inf?
It means that none of your voxel peaks are significant. Since FDR is based on the actual data, it gets set to Inf if nothing passes.
Thank you for your help!
