Hi Alberto,
This is behaviour is expected.
For simplicity, consider cluster extent first: before masking, you may
have a large, significant cluster. When the mask is applied, if the
cluster crosses the mask border, it is inevitably broken into 2 or more
pieces. What remains within the mask may not be large enough to reach
significance.
TFCE comingles, in the same statistic, height and extent (the precise
definition of "extent" here can be found in the original TFCE paper, but
it follows the same principles of cluster extent as above). When
masking, as the extent is modified, so is the TFCE and its corresponding
p-values. The difference is larger where the extent is modified mostly,
i.e., near the borders of the mask, which is indeed what you observed.
Hope this helps!
All the best,
Anderson
On 04/01/2014 11:49, Alberto Inuggi wrote:
> happy new year to everybody !!!!
>
> I would need some info about the effect of masking a randomise analysis in conjunction with a TFCE correction.
> by masking a map, thus investigating less voxels, I expect more significant results as the multiple comparison is performed over less voxels.
> counter-intuitively (at least for me), when applying a certain mask I loose significant voxels, compared to unmasked analysis, even if the disappearing voxels are contained within the mask. the vanishing voxels are 99% close to the border of the mask. which is the reason? can i do anything to avoid this?
>
> thanks in advance
> regards
> Alberto
|