Seed-based functional correlation -- group-level stats

Dear AFNI experts,

I have a question in generating group-level maps of the functional correlation maps that is highly positively skewed.

I basically followed the steps of simple correlation analysis from seed region time series during task performance (https://afni.nimh.nih.gov/SimAna)

  1. original volumes were cleaned by subtracting baseline/nuisance activity included activity from ventricles, white-matter, signal drift, head-motions.
  2. From the clean data, I ran 3dDeconvolve with with the time seed activity from seed regions with removing signal drift (polort) / head motions.
  3. Based on the correlations / R^2 values, I transformed into correlation Z-maps.
  4. I didn’t apply any bandpass filter on the data.

For each individual participant, I have seed-based Fisher’s z-transformed correlation map. The resulting Z-map ranges about [–0.32, 1.02] or so, but my correlation maps are highly positively skewed consistently all participants – except a few regions (around ventricles) showing negative correlations, I see mostly positive correlations across the whole brain. Because mostly all regions exhibit positive correlations in all participants, I get a significant group map from 3dttest++ on z-maps almost all places in the brain (i.e., group map from 3dttest++ against 0).

It is possible the single-subject correlation Z-map can be just wrong. But if this is just the case that they are really positively skewed (mostly showing positive correlations) – I was wondering how I should go about finding a group-level cluster in the dataset that is not normally distributed around 0.

Thanks in advance!
Best, Michelle

Hi Michelle,

It’s not clear as to why you’re concerned about the high proportion of voxels in the brain are positive than negative.

I was wondering how I should go about finding a group-level cluster in the dataset that is not normally distributed around 0.

The current approach (e.g., 3dClustSim) to handling multiple testing does not assume Gaussianality across the brain, nor does it require that the distribution of the voxels in the brain centers around 0. The null distribution is generated through simulations with some parameters (e.g., spatial correlation, voxel size, etc.) from the original data, but not the original effect values themselves.