Hello,
I am working with beta estimates derived from first level modeling in nilearn. I took these beta maps for a series of subjects and performed 3dttest++ in AFNI to look at differences between groups. I am hoping to apply thresholding to the resulting t-statistic image using uncorrected P < 0.0050 and then perform FDR-corrected cluster-size threshold (p < 0.05).
I recently discovered this link (Appendix A: Cluster Correction — Andy's Brain Book 1.0 documentation) which describes a way to perform this if you have residuals exported from analyses in AFNI. However, it looks like it only determines the autocorrelation of noise among a single subject.
My question is the following: can the derived FDR cluster size from this method (on a single subject) be applied to FDR correct clusters from a group-level analysis?
How does one accomplish fdr-cluster correction at the group level?
Thanks so much everyone!
Hi-
Indeed, estimating the average smoothness of noise/residuals occurs on a per-subject basis. The assumption to be able to do clustering on the group level is that the noise characteristics are approximately constant across all subjects. Therefore, you can take the mean smoothness estimate across all subjects and use that for clustering. This is described, with code, here. Note that such clustering even implicitly implies that smoothness characteristics of noise are constant across the brain.
Note that applying clustering might not be the most important thing within results reporting and can even lead to reproducibility issues if taken too literally. It can be helpful to point out the most significant results, but showing results within the full context help quite helpful, as well:
Taylor PA, Reynolds RC, Calhoun V, Gonzalez-Castillo J, Handwerker DA, Bandettini PA, Mejia AF, Chen G (2023). Highlight Results, Don’t Hide Them: Enhance interpretation, reduce biases and improve reproducibility. Neuroimage 274:120138. doi: 10.1016/j.neuroimage.2023.120138
And on a sidenote, I think technically clustering is an FWE correction approach, not an FDR one. FWE aims to control false positive rates among the full dataset, while FDR tries to do the same among reported positives. In practice in neuroimaging, it is rare for FDR to be calculated on a per-voxel basis, and FDR finds more usage in ROI-based reports.
–pt
Indeed, I would discourage using FDR-type correction on clusters. Using FWER correction is certainly better principled.
If you use cluster-level inference you might want to assess the proportion of active voxels therein.
But overall, voxel-level inference is less ambiguous.
Best,
Bertrand