Hi Giuseppe,

> when I want to get more spatially-specific information about the activations, I tend to raise the statistical
> threshold until I get reasonably spatially-restricted clusters. However, it just happens sometimes that, in
> the same study, you have some contrasts that look good at a reasonable threshold of, say, p=0.001 and
> k=20 voxels, while others contrasts give you such large clusters that you are forced to pick a really
> conservative first-level threshold to split them up (eg, p=1e-07), obviously setting yourself up for lots of
> false negatives.

In the early days statistical power was a more prominent issue in the field. However, it seems now we see more and more situations where the whole brain lights up at a typical significance level (cluster p-value of 0.05). This raises a serious question: Is it really the case most of the brain is actually activated, or something is amiss so that we could not achieve a better specificity?

> Furthermore, it is probably not correct to pick a t-threshold for a contrast and a different one for another... or is it?

That does seem to be a questionable practice. A statistic typically serves the purpose of controlling for false positives. Other than that, the statistic value itself is a dimensionless number and does not have any physical meaning; once a proper threshold is set, its functionality is over.

My question to you is the following. If the brain regions are functionally defined (assuming perfect alignment), there should have some clear boundaries between those regions in terms of brain response magnitude. Instead of focusing on the statistical values, would it make more sense to use the brain response magnitude (e.g., percent signal change value) to separate those regions? I guess imperfect spatial alignment across subjects would foil this line of thinking.

> I guess one alternative is to forsake cluster-thresholding altogether and use FDR although, in my experience,
> the FDR thresholding scheme produces results that are quite a bit more erratic compared to cluster-exitent thresholding.

FDR is typically more conservative than the cluster-based correction method. So it may temporally hide the issue, but does not fundamentally solve the problem.

Gang



Edited 2 time(s). Last edit at 09/02/2014 06:33PM by Gang.