Thanks for helping me think through this. In standard practice, we use 3dFWHMx to estimate the ACF parameters based on the residuals of the first-level “fixed effects” model. If we are, say, using SPM or FSL, we do this for each participant’s first-level model and take the average (or median) ACF parameter to use in the 3dClustSim step. I think I have a tenuous understanding of why this is a good estimate of the spatial smoothness of noise or error in a way that’s useful for producing the null spatial distribution, but as you’ll see, it’s not strong enough for me to reason with confidence about the following situation.
Most of the time, we’re interested in a second level model where the brain data is the outcome, i.e., the dependent variable. But what if we do a whole-brain voxel-wise regression with the brain data as the independent variable? Is the spatial smoothness of the residuals at the first-level still a good estimate for the group-level regression: DV ~ BRAIN? In case it needs to be said, in this case DV will be the same for every voxel.
I guess nobody has done anything to come up with a reasonable approach to handling the situation you described. Considering the fact that the spatial relatedness estimated from individual subjects is a rough approximation in the first place for the group-level model, I would say you can still use the same approximation for your data analysis.
That is very helpful to hear. We are also producing maps from the group-level models estimated on permutations of our data as well to use with the -inset option in 3dClustSim as a baseline for comparison to the simulations using the acf estimated from the first-level data. Hopefully this confirms that intuition!
I would like to follow up to ensure I’m doing this correctly. For my model, I’m permuting the data (I won’t get into the specifics here) and generating one whole-brain Z-statistic map per permutation. As a start, I’ve generated maps from 2000 permutations, and then 3dTcat’d them into a 4d nifti. I’m then running 3dClustSim. For example:
Based on reading the docs, I’m assuming it thresholds these maps at .005 and .001 and produces the cluster-extent thresholds to limit false-positive clusters to a proportion less than the various values of athr. I think this is what I want as an empirically-derived FWE correction. Does this sound right to you?
Yes, I am using the -mask option though I forgot to include it in the original email. Thank you for making sure!
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.