You are wrong in assuming that RFT is the same as Bonferroni correction with the number of Resels. This is just a rough approximation to give an idea of how things work.
In practice though RFT accounts for the size but also the shape (surface and radius) of the search space. This means that 2 volumes with the same voxel count (and smoothness) but different shapes (think of a spherical vs elongated volume) will have *different* significance t-threshold (higher for the elongated than the spherical volume).
De: "chenhf_uestc" <
[log in to unmask]>
À: [log in to unmask]Envoyé: Jeudi 22 Octobre 2015 10:53:52
Objet: [SPM] Question about the FWE correction at the peak level in SPM
Dear SPM experts,
I have a question about the FWE correction at the peak level. Actually, the FWE correction is the improved version of Bonferroni correction. If we want to use Bonferroni correction and there is 20000 voxels, so we should let each voxel's p value < 0.05/20000; However, the functional imaging data usually have some spatial correlation, and thus the statistical test of each voxel is not independent. SPM introduce the random field theory to estimate the total number of independent tests. Is this number equal to the resel's number (resolution element)?
In the attached picture 1, there are 1534518 voxels which is equal to 1848.8 resels. Could I understand it in this way? There are 1848.8 independent test in the whole brain statistical analysis. So, the p value in each voxel should be < 0.05/1848.8 rather than 0.05/1534518. If so, in this way, the p=0.05/1848.8=2.7045e-05. Then, because the degree of freedom in this study is 185 (attached figure), I can use the tinv function in the Matlab to calculate the t value: t = tinv(1-2.7045e-05,185) = 4.1336.
I think this value (4.1336) corresponds to the t value of FWE p<0.05 at the peak level. However, in the attached figure, the t value of FWE p<0.05 is 4.985 (FWEp, the left bottom of the figure).
Where am I wrong?
Any help will be greatly appreciate!
Best,
Feng
