Hi Yuzhang,
The thresholding in 3dClustSim is done using z-scores
(so that applying a blur is a reasonable operation,
which is not the case for F-stats, for example).
If there is a reasonable blur to the data and if the
p-value is at least "modestly significant" (is that an
oxymoron?), then bi-sided clustering will commonly not
have an effect. That is because positive and negative
results would tend not to cluster together (assuming
a signed threshold to begin with).
If you are thresholding based on the F-stat, then the
sign is not even relevant, and there will never be a
difference between bi-sided and 2-sided clustering.
- rick