3dclustsim question on ACF option

AFNI Message Board
1

Hello AFNI experts,

I want to double check that I am using the new option of AFNI Clustsim correctly. I have in the past used the version that uses the FWHM and I am surprised by how much different the cluster sizes when using the ACF option.

For example: NN3 2 sided, at pthr 0.05 and alpha 0.05, the original version cluster size is 768 voxels, whereas in ACF version it is 3590. This is a 4 fold increase in voxels needed for a cluster to survive correction. Is this in accordance with what is expected with the ACF option?

I am running cluster correction for a 3dMVM model with both within subject (3 conditions) and between subject regressors. Here is my code for calculating smoothness and running clustim with the ACF option. I should note that my single-subject data was processed in FSL, so the res4d.nii file for single subject data should be equivalent to errts file generated by afni_proc.py.

calculate smoothness from single-subject residuals for each condition of task

foreach n (…)
foreach cope(1 2 3)
3dFWHMx -dset res4d.nii.gz -automask -detrend -acf >> smooth_acf.1D

concatenate smoothness across subjects into one file

cat smooth_acf.1D >> /Volumes/danl/afni_clustsim/clusterize_step1_output_smoothness_residuals_acf.1D
end
end

calculate average across all subjects/ conditions for smoothness ACF parameters.

run 3dclustsim

3dClustSim -mask mask.nii -acf 0.55 5.44 18.0 -LOTS -iter 10000 -nodec -NN 123 -prefix 3dClustsim_ACF_3dMVM_GroupxAgexEmotion

2

Michelle,

What you’re seeing basically confirms what has been reported in the literature: 1) there is some significant amount of spatial heterogeneity, and 2) the Gaussian assumption about the spatial structure in the residuals is not valid. The consequence of these two facts is the liberal control for false positives when using the conventional FWE correction methods including the option -fwhm with 3dClustSim with relatively large p-value thresholds such as 0.05 and 0.01. If you try a low p-threshold such as 0.001, the difference between FWHM and ACF would be much smaller.