---------- Forwarded message ---------- Date: Mon, 30 Jun 2008 16:23:54 +0100 From: Thomas Nichols nichols@fmrib.ox.ac.uk To: jorge luis jbernal0019@yahoo.es Cc: freesurfer@nmr.mgh.harvard.edu, Bruce Fischl fischl@nmr.mgh.harvard.edu Subject: Re: About cluster thresholding
Dear Jorge,
When using a permutation procedure, there's no restriction on what initial threshold you use to define clusters (parametric, random field theory, on the other hand, requires a sufficiently high threshold, like P=0.01 or more stringent).
As for the number of permutations, see http://www.fmrib.ox.ac.uk/fsl/randomise/ and look for the table of 'Confidence Limits'. There you'll see that the confidence bounds on your P-values depend on nPerms^(-1/2) and the true P-value. For a true P-value of 0.05, you need 5,000 to bring the limits down, and 10,000 to ensure that there is less than a 10% margin of error on a true 0.05 P-value.
A couple of caveats, though: The confidence limits / margin of errors assume there is an infinite number of possible permutations, or a very large number. If there are only, say, 1024 possible permutations, there's nothing to be gained from running more (there are no more to run)! Also, since 5,000 or 10,000 will take a long time, I always recommend first running with, say, 100 or 500, just to make sure you've got the right data, have a decent cluster-forming threshold, etc, before running a 'gold standard' number of permutations.
-Tom
PS: I have to plug my recent work with Steve Smith, on avoiding the problem of 'Which cluster-forming threshold?', with a method called Threshold-Free Cluster Enhancement; see http://www.fmrib.ox.ac.uk/analysis/techrep/tr08ss1/tr08ss1.pdf and http://dx.doi.org/10.1016/j.neuroimage.2008.03.061 . It's currently implemented in FSL's randomise, though is a bit slow. An immenent update to FSL will speed it up considerably.
On Mon, Jun 30, 2008 at 4:08 PM, Bruce Fischl fischl@nmr.mgh.harvard.edu wrote:
Hi Jorge,
I'll leave this for either Doug or Tom Nichols, who understand this much better than me. I would guess that 5000 is not enough, but I could be wrong.
cheers, Bruce On Mon, 30 Jun 2008, jorge luis wrote:
Thank you very much Bruce
I have another question:
I plan to use permutation testing in order to correct for multiple comparisons.
Is it enough to perform 5000 iterations for an initial voxel_wise threshold of 0.005?
Is there also any heuristic to select this arbitrary initial threshold and theoretical results for the required number of iterations for the permutations?
I appreciate very much your help
Jorge Phd Student
--- El dom, 29/6/08, Bruce Fischl fischl@nmr.mgh.harvard.edu escribió:
De: Bruce Fischl fischl@nmr.mgh.harvard.edu
Asunto: Re: [Freesurfer] About the area of the vertices in the surfaces Para: "jorge luis" jbernal0019@yahoo.es CC: freesurfer@nmr.mgh.harvard.edu Fecha: domingo, 29 junio, 2008 5:22 Hi Jorge,
yes, the area of a vertex is 1/3 of the area of all triangles that it is a member of, and it should sum to the area of the surface. The average subject has a correction factor that accounts for the "lost" area due to averaging, so you'll need to apply this if you want it to be in true mm (in the average sense across your subjects). mris_info ~/local_subjects/fsaverage/surf/lh.white reading group avg surface area 864 cm^2 from file Reading in average area /homes/4/fischl/local_subjects/fsaverage/surf/lh.white.avg.area.mgh SURFACE INFO ======================================== type : MRIS_TRIANGULAR_SURFACE=MRIS_ICO_SURFACE num vertices: 163842 num faces : 327680 num strips : 0 surface area: 70394.8 AvgVtxArea 0.429650 AvgVtxDist 0.757401 StdVtxDist 0.227951 group avg surface area: 86444.7 . . .
thus, you would need to multiply your areas by 86444.7/70394.8 to get mm in the average cross-subject sense. cheers, Bruce
On Sun, 29 Jun 2008, jorge luis wrote:
Hello all
I have some questions:
I would like to know how freesurfer computes the
vertexÿÿs wise areas? eg. thouse in lh.area.
It should be expected that the sum of the areas of all
the vertices across a surface equals the total surface area?
To be much specific: Can I compute the area of an 'activated'
cortical ROI in the average subject by simply adding the area of all its vertices?
In adavance thank you
Jorge Phd Student
¿No te gusta tu dirección de correo?Consigue una que te guste de verdad - millones de
direcciones de correo disponibles en Yahoo!
http://es.docs.yahoo.com/mail/nueva_direccion.html
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
______________________________________________Enviado desde Correo Yahoo! La bandeja de entrada más inteligente.