I have been working for a while now trying to get cluster thresholds by using 3dClustSim for a 3dLME group analysis I ran, but the thresholds I am getting still seem strangely low. For a wholebrain with functional data smoothed at 6 with 3dBlurToFWHM, at p threshold of .005, alpha of .05, NN2, bi-sided, it recommends a threshold of 16.3 (voxel size 3.125 mm x 3.125 mm x 3.125 mm). I am calculating my ACF values using 3dFWHMx and the input file is the residuals I collected from the 3dLME analysis I ran (which was supported as an option here: https://afni.nimh.nih.gov/afni/community/board/read.php?1,156149,156170). Below are the ACF values output:
0 0 0 0
0.572146 2.46819 7.35051 6.73157
Do you have any ideas about what I might be missing? Please know if I can add additional information to help get to the bottom of this.
After pondering for a while, those numbers do not seem out of the realm of reasonableness. Your blur level is pretty low: 3.125 mm voxels, with a final blur (not an added blur) of FWHM = 6 mm.
That would be akin to adding a blur of FWHM = 3 mm, or so, which is very small.
So if the final FWHM is supposed to be 6 mm (and that is estimated with a standard, Gaussian ACF), you are getting a mixed-model ACF FWHM estimate of 6.7 mm, which seems to be in the correct ballpark.
To be sure, note what the 4 parameters are. The first is the fraction of the Gaussian term in the ACF model (so the linear fraction is .43 = 1-.57). The Gaussian term is 2.5 mm (small), and the linear term is 7.35. But that leads to a reasonable FWHM of 6.73 mm.
Also to be sure, are you applying a brain mask in this computation?
Hi! I thought of creating a different topic, but I think my doubt is related enough to this one, so I'll recycle it.
I was also trying to get a cluster-level thresholded image from a group obtained with 3dLMEr. But I am not 100% sure on how to do it. My idea is to extract the thresholded image that would result from applying a cluster FDR correction (for an alpha of 0.05) to the voxel-level threshold of p<0.001.
I was thinking on the approach that is described in this topic: using the residuals from 3dLMEr to input the acf values to 3dClustSim. Like this (note that I'm not using a mask because the image was already masked during 3dDeconvolve):
3dFWHMx -input residuals_lme -acf
3dClustSim -acf a b c -athr 0.05 -pthr 0.001
From the function help and the papers on multiple comparisons corrections, I guess that this is the most accurate method, but I am not sure if I should proceed with the group residuals. Mainly because I am not sure I understand the implications of doing this at the subject level, as compared to the group level.
Also, I know this approach tells you, in a nutshell, how big your cluster must be to be considered significant (according to the specified criteria). But how could I get the thresholded image without the GUI? Just 3dcalc?
If you could provide some feedback I would be eternally grateful
The accuracy of handling analytical multiplicity is likely in the eye of the beholder. In my opinion, three major concerns arise from the voxel-level p-threshold of 0.001, which can result in substantial reproducibility challenges. First, the conventional cluster-based approach implicitly assumes a uniform brain distribution across the brain, a premise that is substantially violated. Second, the resulting penalty is overly harsh and lacks adaptability. Lastly, there is an exclusive emphasis on identifying "false positives," while "false negatives" are totally overlooked.
As a recommended alternative, consider adopting a "highlight, but don’t hide" strategy, as exemplified here.
Gang Chen
1 Like
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.
