Hi Tom,
Thanks for the reference and explanation. Would you also recommend, since I am performing a ROI analysis, doing an uncorrected test, at P=0.05 or 0.01 for the regions I am interested in?
regards
Nelleke
________________________________
From: FSL - FMRIB's Software Library on behalf of Thomas Nichols
Sent: Fri 7/4/2008 7:43 AM
To: [log in to unmask]
Subject: Re: [FSL] thresholding question
Hi folks,
Saad's right, as the threshold drop's the expected EC becomes a worse approximation of the expected number of clusters. There's yet another problem, though: For high thresholds, the null distribution of cluster size is approximately exponential. As the threshold drops this approximation gets worse, and cluster P-values get less accurate.
For examination of RFT cluster size performance as a function of threshold and smoothness, see: S. Hayasaka and T.E. Nichols. Validating cluster size inference: random field and permutation methods. NeuroImage, 20:2343-2356, 2003.
On Thu, Jul 3, 2008 at 2:22 PM, Saad Jbabdi <[log in to unmask]> wrote:
My -little- understanding of GRF is that:
The rationale behind GRF is that the -expected- euler characteristic (EC) approximates the FWE probability.
EC basically counts the number of peaks minus the number of holes (after thresholding). If your threshold is too low, you will have many holes, and the EC will not represent the number of peaks, which you want in order to approximate the FWE..
Saad.
On 3 Jul 2008, at 13:52, Wouwe, Nelleke van wrote:
Hi Tim,
I am not sure if I understand you correctly: are you saying that I
cannot use Z-values below 2 because GRF theory is not valid anymore? Why
is that?
Thanks for your help
Nelleke
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
Behalf Of Tim Behrens
Sent: 03 July 2008 14:20
To: [log in to unmask]
Subject: Re: [FSL] thresholding question
Hi - if you are using Gaussian Random Field theory for your inference
(i.e. if you are using Flame and not randomise) then it breaks down at
low Z values (less than about 2).
Cheers
T
On 3 Jul 2008, at 12:56, Nelleke Van Wouwe wrote:
Hi all,
From some previous discussions and Matt Brett's page on cluster
thresholding I understand that you chose the z value based on the
activation you are looking for;high z for small regions and low Z for
big regions (given p<0.05, corrected).
I performed a ROI analysis (pre-masked the data with part of the
fusiform gyrus and parahippocampal gyrus)with a (corrected) cluster
significance threshold of P=0.05 and I tried a very low Z threshold of
1.1.
According to the clusterlist below I find a large cluster (232
voxels) that
is significantly active (P<0.05) with z-max 1.9.
I wonder if there are any rules of thumb to determine an appropriate Z
threshold (other than the default) and would this z = 1.1 still be a
reasonable threshold?
Thanks!
Nelleke
Cluster List
Cluster Index Voxels P -log10(P) Z-MAX Z-MAX X (mm) Z-MAX Y (mm) Z-
MAX Z
(mm) Z-COG X (mm) Z-COG Y (mm) Z-COG Z (mm) COPE-MAX COPE-MAX X
(mm) COPE-
MAX Y (mm) COPE-MAX Z (mm) COPE-MEAN
5 232 0.0356 1.45 1.83 42 -50 -22 42.7 -47.8 -24.2 193 42 -42 -30 149
4 41 0.039 1.41 2.26 -44 -62 -26 -42.7 -60.2 -25 243 -44 -62 -26 163
3 26 0.0394 1.4 1.68 34 -34 -20 31.5 -35 -19 134 34 -34 -20 107
2 24 0.0394 1.4 1.45 -32 -38 -20 -30.8 -38.3 -18.6 103 -32 -38 -20
87.6
1 1 0.0405 1.39 1.15 24 -48 -6 24 -48 -6 110 24 -48 -6 110
**********************************************************************
This email and any files transmitted with it are confidential and
intended solely for the use of the individual or entity to whom they
are addressed. If you have received this email in error please notify
the system manager.
**********************************************************************
---------------------------------------------------------------------------
Saad Jbabdi,
Postdoctoral Research Assistant,
Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222545 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~saad
---------------------------------------------------------------------------
--
____________________________________________
Thomas Nichols, PhD
Director, Modelling & Genetics
GlaxoSmithKline Clinical Imaging Centre
Senior Research Fellow
Oxford University FMRIB Centre
|