10 May
2011
10 May
'11
9:36 a.m.
This email to the FSL mailserver list may be of interest to local YNiC
FSL users
Gary
----------------------------------
Dear David,
The short answer is, no, there is no way to justify Z=1.7 as a
cluster-forming threshold in FEAT.
The problem is that the cluster size P-values are based on Random Field
Theory (RFT), and RFT makes various approximation that are only valid
for high thresholds. While one early reference (Petersson et al, 1999)
specified cluster-forming threshold P=0.01 / Z=2.33 as a practical lower
limit for accurate results, later work (Hayasaka & Nichols, 2003; Silver
et al, 2010; K. Worsley personal communication) found even that level
was unstable, and instead recommended P=0.001 / Z=3.09 as a lower limit
on a cluster-forming threshold to ensure accurate inferences.
On the other hand, if you are using randomise, you can safely use any
cluster-forming threshold (though if you go too low, the clusters might
be too extended and spindly to be interpretable). If using randomise,
though, check out TFCE as away to avoid specifying any particular
cluster-forming threshold.
To answer your second question, the cluster-forming threshold used at
lower-levels is only used to make inferences *at* the lower level, and
is ignored at higher levels. Only the cluster-forming
threshold specified in the top-level FEAT analysis matters for the
top-level results.
-Tom
Petersson, K. M., Nichols, T. E., Poline, J.-B., & Holmes, A. P. (1999).
Statistical limitations in functional neuroimaging II. Signal detection
and statistical inference. /Phil. Trans. R. Soc. Lond. B/, /354/, 1261-1281.
Hayasaka, S., & Nichols, T. (2003). Validating cluster size inference:
random field and permutation methods. /NeuroImage/, /20/, 2343-2356.
Silver, M., Montana, G., & Nichols, T. E. (2010). False positives in
neuroimaging genetics using voxel-based morphometry data. /NeuroImage/.
Elsevier Inc. doi: 10.1016/j.neuroimage.2010.08.049.
On Mon, May 9, 2011 at 1:06 PM, David Soto <d.soto.b(a)gmail.com
<mailto:d.soto.b@gmail.com>> wrote:
Hello,
I am finding that when I use a Z<1.7 to define the cluster size at
he highest level analyses, I am getting some interesting activations
in regions that I don't see when I use Z<2.3. In both cases I use
cluster thresholding and p<0.05 whole brain corrected.....
I know the Z value used to define cluster size prior to correction
for multiple comparisons is arbitrary, but is it there any paper
that I can use to justify
in my study why a Z<1.7 was used instead of Z<2.3?
Would it be right to say that poststat results with a Z<1.7 are more
lenient than with a Z<2.3? I feel it this is not necessarily right
but can you please advise?
A second question I have is about poststats as implemented in FEAT....
Say that I have done a lower level analyses - for session, a 2nd
level -across session within subjects- and 3rd level -across subjects-
do I need to set up the Z<1.7 at the first and second level and
third level analyses - or does FEAT bring the unthresholded data
from lower level analyses to the higher levels and then use the Z
score specified at the highest level analyses? in other words, will
I get the same results if I specify
Z<1.7 across all levels or whether I do Z<2.3 for first and second
level and Z<1.7 at the higher level?
Many thanks,
David
--
____________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick
Coventry CV4 7AL
United Kingdom
Email: t.e.nichols(a)warwick.ac.uk <mailto:t.e.nichols@warwick.ac.uk>
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax: +44 24 7652 4532