Hi Torsten,

First, note that using your inputs (2 mm^3 with 11 mm FWHM and a
mask of 180000 voxels), I get basically the same results as you,
300 voxels are required. This is matching an uncorrected p of
0.005 (as was used in Rameson) with a corrected p of 0.05 (as is
common).

As to how you and Rameson differ, let me just make a few comments.

1. Using 3 mm^3 (as they did) gives a result of 84 voxels. Note
that 84 voxels at 3 mm^3 is almost the same volume as 300 voxels
at 2 mm^3.

2. Their mask could have smaller.

3. They did not seem to report the estimated FWHM, which they
should have. So they possibly used only 8 mm for it, which
would not be appropriate.

4. Using a blur of only 8 mm gives a cluster size of 44 voxels,
still more than twice what they reported.

5. They do not even say what their corrected p-value is. But you
can be sure that "balancing Type I and Type II errors" means it
is probably higher than 0.05.

6. They talk about an FDR correction with these values, presumably
of the clusters, not the voxels.

7. A threshold of 20 voxels does not even show up on the p<0.005
line, though it would be appropriate at p<0.001.


In short, there are a few main differences (your voxels are smaller,
your blur is higher, and you might actually correct at 0.05), or
whatever you choose. Their correction seems a bit gentle.

- rick