Hello, this is my first time using all the updated 3dClustSim options and wanted to clarify some things. I am using 3dClustSim to correct the results of 3dMEMA. In the past, my 3dClustSim command was the following:
3dClustSim -mask GM3+tlrc -fwhmxyz 8 8 8 -pthr .025 -athr .05 -NN 2 -prefix output
(GM is just a mask of only the gray matter voxels)
I understand that the NN is no longer needed but for the new 3dClustSim,
1.Does the pthr still need to be halved? So in the past I set the pthr to .025 to get the voxelwise threshold of .05. Is this still the same for the more updated use?
- If you want to consider both A-B and B-A you would use the two sided output correct?
(1) I no longer recommend the use of the -fwhm or -fwhmxyz options. Instead, I recommend that you use 3dFWHMx -acf to get the mixed Gaussian+exponential model fit parameters (a,b,c) from your subject level residuals, take the mean of those across subjects, and use those as input to 3dClustSim with the -acf option. If you are using an up-to-date version of afni_proc.py for your individual subject analyses, then it will give you the acf parameters now.
(2) You do not halve the p-values. The 1-sided and 2-sided testing inside 3dClustSim now does that for you, and gives you the tables for each case of sided-ness and NN connectivity-ness. You probably should just use the -LOTS option (along with the -prefix option) to get the values you want in nicely formatted files.
(3) This stuff is the subject of a paper we are submitting very soon (hopefully, Friday). When we can post a link to it, it will go here, never fear.
Could you advise me how to use the cluster-level corrected alpha of .05 to report the group results now? Or could you let me know your new paper - does it address this issue? Is this method still valid: " the alpha-level for voxelwise statistical analysis was determined by simulations using the 3dClustSim program of the AFNI toolkit. For these simulations, the smoothness of the data in three directions was estimated using 3dFWHMx on the residual time series of gray-matter voxels in each participant and then averaged across participants (FWHMx = xxx mm; FWHMy = xxx mm; FWHMz = xxx mm). Based on a voxel-level uncorrected alpha of .001, simulations indicated a minimum cluster extent of xx voxels for a cluster-level corrected alpha of .05".
Adding to this thread re. smoothness estimates using first level average vs group residuals, in the Eklund et al. (2016 PNAS) paper, the authors note that the group smoothness used by 3dClustSim may be too low (compared with SPM and FSL), because (unlike SPM and FSL) AFNI averages smoothness estimates from the first-level analysis.
In light of this, would it not be more preferable to use residuals form the group analysis to estimate the smoothness?
I appreciate your help clarifying this.
It would probably make sense to take a look through these articles for descriptions of where clustering is at in AFNI-- both regarding suppositions made in the Eklund et al. paper and with new methods that Bob has cooked up:
There is an “executive summary” of the longer article, which might be extra useful as a start.
Current (in this everchanging world) processing recommendations are given at the end of the long one.