Hi All,
I've searched the board for a discussion on this point but couldn't find one -- excuse me if this has been hashed out.
I am analyzing a contrast between two experimental conditions and using Alphasim to find reliable clusters. When the single voxel threshold is set relatively high, say p < .0001, reliable clusters are at least 4 voxels large, and there are a few of those. When the single voxel threshold is set to p < .01, reliable clusters are at least 50 voxels large, and there are a few of those as well. (Both analyses equally control for whole-brain alpha at p < .05)
Now, in certain regions the two types of clusters overlap (i.e., the small clusters fall within the larger ones), but in others they don't. I feel that choosing one analysis over is a mistake as this provides a skewed view of the data.
But, if I report both, should I control for "multiple tests"? I theory, I could have systematically iterated through 20 single voxel thresholds and examined the results.
Related, some permutation methods use "cluster mass" as a statistic rather than cluster size. That is, from the simulations they construct a distribution of a statistic that reflects [cluster size * mean_voxel_intensity_in_cluster] (say for all clusters greater than 3 voxels at voxel threshold p < .05). With cluster mass, small clusters that are highly reliable have the same mass as larger clusters where voxels are less reliable, and so such statistic could simultaneously select for both sorts of clusters that I find using alphasim.
Any possibility of extracting this statistic from Alphasim? I suppose 3dclust could then be used to extract clusters whose mass is above the threshold.
Cheers,
Uri