Dear Stephen,

Thank you, that is very helpful. Is there any way to view a voxel level FDR correction of the results in SPM or fsleyes?

I’m a bit uncomfortable with uncorrected voxel level threshold.

Any references would also help me out, I appreciate it!

Sincerely
Andreas Voldstad

* SPM Random Field Theory RFT is about doing topological inference, in other word cluster inference or feature level inference! Voxel-wise inference (FWE and FDR) is not enough as the signal is continuous, we cannot just say that a voxel is activated or not, we must account for the underlying signal! But we can do both: correction at voxel-level AND at cluster-level/feature-level. Hence cluster-mass FDR is as valid as cluster-mass FWE! This is just another way to control the number of false positive clusters (whereas voxel-wise FDR controls for the number of false positive voxels). To understand the problem, here is an example: a SPM with P-FDR threshold with 95 voxels in one cluster and 5 spurious voxels in 5 distinct places, it has 95% FDP rate voxel-wise, but only 16,7% cluster-wise because of 6 clusters found only one was correct! Thus regions found with voxel-wise FDR are >80% false positive, only voxels are <5% false positive, so we need to also control cluster-wise FDR to get <5% false positive regions ! Chumbley, Justin R., and Karl J. Friston. "False discovery rate revisited: FDR and topological inference using Gaussian random fields." Neuroimage 44.1 (2009): 62-70. and also Chumbley, J., Worsley, K., Flandin, G., & Friston, K. (2010). Topological FDR for neuroimaging. Neuroimage, 49(4), 3057-3064.
* Thus paradoxically, although voxel-wise FDR inference maximizes voxel spatial specificity, it lacks regional specificity! And inversely for cluster-wise inference! Thus, cluster-wise/topological inference might be better for regional studies! But then don't know what region exactly activated, only that there is a signal somewhere in that area! It depends if data is smoothed, then cluster-wise is better. Problem is distributivity aka dependence between voxels, with cluster level on rft we work again on discrete features! Also for conjunction either do same voxel-wise threshold, intersect and do clusterwise if interested in regional changes, or do min stat maps then do voxel + cluster correction! Alternative is to evaluate manually: report only the top regions with lots of voxels, not the spurious small ones!
* Comparison of thresholds families summary: peak-FWE < peak-FDR < cluster-FDR < voxel-FDR, thus peak-FDR and cluster-FDR are more conservative than voxel-FDR! Peak-FDR and cluster-FDR are both topological correction, not direct features of the signal (contrary to voxels). (Chumbley, J., Worsley, K., Flandin, G., & Friston, K. (2010). Topological FDR for neuroimaging. Neuroimage, 49(4), 3057-3064). And according to Eklund, FDR voxel-wise is worse in terms of reliability than FDR cluster-wise: «no correction < FDR voxel-wise < FDR cluster-wise < FWE cluster-wise < FWE voxel-wise» https://blogs.warwick.ac.uk/nichols/entr...
* If you want both perfect regional and voxel-wise spatial specificity, then use voxel-wise FWE. But the price to pay is a very low sensitivity!
* Topological (RFT) inference methods: peak < cluster < set in terms of sensitivity, and inversely for spatial specificity. They can all be reported if in a nested, step down fashion. Toga, A. W. (2015). Brain mapping: An encyclopedic reference. Academic Press.
* Smoothing totally messes up voxel fdr (because it creates voxel dependency), even for voxel fdp rate!! up to 50%!! so just use cluster fdr.

* Use cluster-level correction for fMRI, and voxel-level correction for PET (maximizes sensitivity for each). Voxel-level correction is a subcase of cluster-level correction when k = 0. Friston, K. J., Holmes, A., Poline, J. B., Price, C. J., & Frith, C. D. (1996). Detecting activations in PET and fMRI: levels of inference and power. Neuroimage, 4(3), 223-235.

* Prefer to use cluster-mass (or better its generalization TFCE) + at least 800 iterations, they are validated and corrects for multiple comparison correctly! Pernet, C. R., Latinus, M., Nichols, T. E., & Rousselet, G. A. (2015). Cluster-based computational methods for mass univariate analyses of event-related brain potentials/fields: A simulation study. Journal of neuroscience methods, 250, 85-93.

Originally posted by Andreas Voldstad:

Actually, I just figured that one out on my own!

Sincerely,

Andreas