Hi Giuseppe,

Thanks for the intriguing question. Here are just some random thoughts of mine about the issue.

> since those maps are typically thresholded using a combination of voxel-wise and cluster-extent
> thresholding, I am not so sure that the region of overlap of the thresholded maps can be
> considered statistically significant.

From the pure statistical perspective, I think that those overlapping regions can be legitimately considered statistically significant in the sense that the false positives are properly controlled at some nominal level (e.g., 0.05 at the cluster level) if all the assumptions about the models hold well, unless there are some flaws in the analysis pipeline. For example, suboptimal alignment across time and across subjects would have impact on regional specificity, but that is not a statistical issue.

> As discussed at length in a recent paper (Woo, Krishnan & Wager, 2014, Neuroimage, 91:312-419),
> when using cluster-extent based thresholding we cannot make any inference about the activation of
> specific locations within a suprathreshold cluster: we can only say that "*some* signal must be
> present *somewhere* in the cluster" (quote from the above paper). Consequently, if the location of
> "true" activation within a given cluster is undetermined, it seems to me that the intersection area of
> two statistical maps that have been thresholded with the cluster-extent method cannot be claimed
> to be itself a spatially-precise locus of significant activation.
>
> Is this a legitimate concern, and if so, what would be an alternative procedure for conjunction analysis?

It's a legitimate concern about the issue of anatomical specificity. A couple of comments:

1) We know that a liberal uncorrected p-value demands a large cluster threshold while a stringent voxel-wise p-value leads to a relatively small cluster threshold. However, there is some wiggle room for specificity improvement even though the room might be not as large as desired. For example, instead of specifying some default range (e.g., 0.001 to 0.01) for the voxel-wise p-threshold, one could vary the uncorrected p-value in a wider interval when running 3dClustSim to obtain an optimal counterbalance between specificity and cluster survivability.

2) Even though smoothing is supposed to improve signal-to-noise ratio and to reduce the effect of misalignment, over-smoothing may contribute to the specificity problem. Sometimes people tend to overemphasize the statistical significance while paying little attention to some extra useful information. Such auxiliary information can improve or ameliorate the specificity problem. For example, effect amplitude (BOLD response strength) may provide an extra layer of filtering in clusterization, but this is currently not well practiced. In addition, prior anatomical structural information and better spatial alignment tools may help.

3) Part of the problem may be caused by an ill-posed model. For example, Pearson correlation is relatively crude when modeling the connectivity among the brain regions considering the following: (a) It measures the overall synchronization over the whole time period between the two regions, but the synchronization may vary from time to time; (b) No delay is considered when measuring synchronization, and neither is the variability of hemodynamic response shape.

Gang



Edited 3 time(s). Last edit at 08/29/2014 09:33AM by Gang.