There are two intriguing questions here:

1) Can one choose a relatively high voxel-wise p-threshold value (e.g., 0.1) as long the cluster-level significance level alpha) is controlled at 0.05?

Theoretically it should be fine because it's the cluster-level significance that matters more. A voxel-level p-value of 0.1 can be considered marginally significant. And if you adopt one-tailed thresholding that is prevalent in the field, a two-tailed p of 0.1 is basically a one-tailed p of 0.05 (assuming t-statistic).

2) At each preset voxel-wise p-threshold value (e.g., 0.005) only some number of clusters could survive the cluster-level significance thresholding. Due to the anatomical constraints, not all brain regions would be able to fit a Procrustean cluster size. That is, big regions would be able to reach to the final cluster list with a relatively high voxel-wise p-value, while small regions would have to work much harder (with very low voxel-wise p-value) to earn a surviving spot at cluster level. Is it reasonable or legitimate to adopt a different pair of voxel- and cluster-level thresholds for each region (from the output table of 3dClustSim)? I don't have a definite answer for this.

Gang



Edited 2 time(s). Last edit at 07/22/2013 02:25PM by Gang.