3dClustSim: a non-intimidating guide

Dear Colleagues

I have an absolute beginner question (I`ve just installed AFNI for the first time…):

how to run the 3dClustSim to determine the cluster size threshold for 2nd level contrasts? Results come from an ICA analysis in GIFT (spatial maps), preprocessing was done with SPM12. Most importantly: what is the INPUT for both 3dFWHMx and 3dClustSim? Is it an actual t-contrast image (e.g. thresholded at 0.05 uncorrected)?

I read previous messages, but still I`m very unsure how to do that.

Your answer is much appreciated.

Your question may sound simple, but it is not – at least, the answer is not simple.

3dClustSim is based on simulating the noise field that interferes with detection of the “true” signal in the dataset. To do this simulation, the program needs some statistics about the the spatial smoothness of the noise. For the case of FMRI time series analyses, these statistics come from analyzing the residuals left over after subtracting the model fit at each brain voxel – the residuals are our best stand-in for the noise.

But in your case, I have no idea what you can use as a stand-in for the noise. You can’t really analyze the noise statistics from a map that contains signal mixed with noise, since you would get measurements of the signal spatial smoothness melded with the noise spatial smoothness.

It is this concept of analyzing the spatial statistics of the noise that people find confusing. It’s not a common idea except in imaging, and most people learn statistical analysis ideas in applications where spatial structure in the noise isn’t even a concept (e.g., genetics).

In short, I don’t know how to help you, and in fact I believe that 3dClustSim is not a proper tool for your problem.

And in general, I think it is a bad idea for someone who does their entire analysis, including group statistics, in FSL or SPM or any other package, to turn to 3dClustSim to validate their results.

In the near future, I am going to disable the Gaussian spatial autocorrelation function (ACF) option in our clustering software, and only allow the use of the more complex “mixed” ACF model, which is usually more conservative. Even then, I would only recommend use of our model for cluster thresholding when the group analysis is done with one of our programs (e.g., 3dttest++). Any other use 3dClustSim even with the ACF option strikes me as getting close to “shopping” for results.


Thank you kindly for your comments. I get and appreciate your point. This is also my observation that sometimes we deal with misunderstanding and misusing of methods in computational neuroscience. Nevertheless, I could see a significant amount of publications, even in reputable journals applying the 3dClustSim for thresholding cross-platform. I wonder, why should it necessarily be bad, if it is done correctly. I have to admit, I find the 3dClustSim approach to thresholding very elegant.

Best wishes