I have a question concerning the correction for multiple comparisons for VBM data. I have processed my VBM data with the cat 12 toolbox for Matlab, and then run t-tests between my groups of interest in AFNI using 3dttest++. In order to correct for multiple comparisons the results of 3dttest++, I followed the procedure that is now recommended, including getting the data smoothness with 3dFWHMx, and then entering the ACF values into the 3dClustSim command. For p=0.01 at alpha=0.05, however, the results of 3dClustSim give me a cluster size of over 1550 voxels, which appears way too strict compared to the size of the regions that I would expect to differ between the groups.
I was wondering if you could confirm whether the use of 3dFWHMx and 3d ClustSim is indeed appropriate to analyze VBM data, and let me know if there is any other way of correcting for multiple comparisons for this kind of data.
Thank you very much in advance.
I don’t see a problem with the general methodology of using 3dFWHMx and 3dClustSim (with -acf). The evaluation of the impact of -acf within those tools (3dFWHMx and 3dClustSim) has been focused on the functional data, and while the Eklund paper does say that anatomical data shared similar problems to functional data, I’m not sure if the problem exists to the same extent. For what it’s worth, Douglas Greve has said he’s doing simulations on Freesurfer output and the results are perhaps not as extreme as the functional results.
The likely reason for such large cluster sizes is that your smoothing kernel is large, having done a similar analysis with the VBM toolbox years ago, I noticed that the change of a 6mm kernel to an 8mm kernel can have pretty substantial impacts on the necessary cluster size as your data is in 1mm isotropic voxel sizes. You could reduce your smoothing and see if that helps, and this would likely would reflect your theoretical interests to evaluate VBM differences in smaller brain structures.
Your other options are to use FDR correction, or permutation testing. For permutation, there’s FSL’s Randomise[/url], Eklund’s own [url=https://github.com/wanderine/BROCCOLI]Broccoli[/url] software. and [url=http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic-research/nichols/software/snpm]SnPM.
thank you very much for your reply. I will try with the other options you suggest.
let me know if there is any other way of correcting for multiple comparisons for this kind of data.
Try the permutation approach embedded in 3dttest++ through option -Clustsim, which does not make any assumption about the spatial correlation structure.