Hi Gaurav,
1. There are probably some inhomogeneities in terms of spatial correlation across the brain, and its impact is unknown if one simply smoothes the original EPI data with a uniform kernel size. However, if this is something of concern, you may consider adopting a uniform smoothing approach by using 3dBlurToFWHM, which would adaptively smooth the original data to achieve a uniform correlation across the brain.
2. If you use 3dBlurToFWHM to smooth the data, you don't need to estimate the FWHM values any more because the final correlation values are the parameters you provided for 3dBlurToFWHM.
3.
>the probability of getting a cluster in the average of those 20 volumes
> will be much smaller than the probability of getting a cluster in a single
> volume. Right?
No, not necessarily.
If you only take the effect estimates (beta values) from each subject to group level, you assume the accuracy (or standard error) of each effect estimate is the SAME across all subjects, and the significane at group level depends on how consistent the effect estimates from those subjects are, and has NOTHING to do with the significance at individual level.
However, if you take both the effect estimates (beta values) and their reliability information (embedded in the t-statistic values) from each subject to group level, both within-subject and cross-subjects variability will enter into the significance assessment at group level. It's hard to say which of the two, within-subject or cross-subject variability, would have an upper hand because it depends on the specific dataset.
> Hence, the clustering threshold for a subject is too stringent
> at the group level. Is there a way to accurately estimate the
> cluster threshold at the group level?
You're right that the cluster thresholding at group level is currently based on the spatial correlation estimates at individual subject level. Ideally it would be better to make such assessment based on the residuals of the group model. I don't know how big the deal is, but I agree it's something worth further exploration.
Gang
Edited 1 time(s). Last edit at 10/24/2012 04:00PM by Gang.