Hello Brian and Alecia:
1) The selection of connectivity radius is somewhat arbitrary. For example,
to allow formation of clusters diagonally across voxels, but in-plane, use
rmm = 2.66, i.e., sqrt(1.875^2+1.875^2)= 2.65165. But, again, the choice
is arbitrary.
2) The Pthr is an arbitrary input parameter. There is no "correct" value.
Rather, program AlphaSim can be used to determine the cluster size threshold,
corresponding to the input Pthr, which yields the desired overall Alpha level.
Hence, by using different Pthr inputs, one can obtain the tradeoff between
the individual voxel probability threshold (Pthr) and the cluster size
threshold (Cthr) for a specified probability of false detection anywhere within
the volume (Alpha). This is illustrated in Figure 3 from file AlphaSim_fig.ps.
The Pthr is neither one-tailed nor two-tailed. It is the individual voxel
probability threshold; i.e., the probability of a false positive for
an individual voxel.
For your particular case, I believe that the .0167 number should be Alpha
(i.e., the output of the program), not Pthr (the input to the program).
3) The p-value associated with the slider bar is one-tailed or two-tailed,
depending on the particular statistic. For correlation coefficients and
t-statistics, it is two-tailed. For F-statistics, it is one-tailed.
4) Whether to use a two-tailed t-value associated with the Pthr, or a
one-tailed t-value depends on whether you are conducting a two-tailed t-test
or a one-tailed t-test.
5) It's possible to create a sub-brick that contains the p-values
associated with the individual voxel t-stats. This can be done using the
3dcalc function fitt_t2p (see 3dcalc -help for more details). This p-value
sub-brick can then be "glued" to the bucket dataset using program 3dbucket.
6) I recommend using a mask as input to AlphaSim. Even for whole brain
clustering, just restricting the number of voxels to those voxels inside the
brain significantly improves the results.
7) For ROI analysis, you can use a mask that restricts consideration to just
those voxels inside the ROI. In this case, I would set rmm to 1000. The reason
is that, with a pre-defined ROI, any active voxel will be counted. In other
words, for ROI analysis, every voxel within the ROI is "next to" every other
voxel within the ROI. Furthermore, the ROI would not necessarily have to be
a connected region. If you use this approach, then the cluster size threshold
can be interpreted as simply a count of the number of voxels within the ROI
that exceed the threshold. Therefore, no actual "clustering" is required.
8) Another way to look at the output from AlphaSim: If you hold Pthr constant,
then the Alpha for each Cthr can be interpreted as the p-value corresponding
to a cluster of size Cthr. That is, instead of thresholding at Alpha, you
can report p-values corresponding to individual clusters of activation based
on the size of the cluster. Similarly, for an ROI, the p-value corresponding
to the ROI is determined by the number of active voxels within the ROI.
9) Important: The output cluster size threshold from AlphaSim is specified in
units of Orig voxels. Now, in order to actually perform the clustering, it is
necessary to specify the cluster volume in mm^3. So, to enter the 3dclust
"vmul" input, one must multiply the cluster threshold from AlphaSim (in # of
voxels) by the volume of an Orig voxel (in mm^3). Since the cluster size
threshold is now specified as a volume, it is independent of the size of an
individual voxel (in Tlrc space).
Doug Ward
