Dear Xiaochu,
I think I understand your question: You're comparing the size of clusters with a 0.001 P threshold (0.999 1-P) on the _vox_tstat image and the size of clusters (not their significance) on the _maxc_tstat image.
The answer: First, a 0.001 significance threshold for 49 DF T image is 3.2651 (by parametric theory), so you might well expect the clusters to be bigger than a raw T image thresholded at 3.5. BUT, more importantly, the _vox_tstat image is an uncorrected *nonparametric* P-value image, where P-values have been determined by permutation at each voxel. An 0.001 threshold on that image *is* *not* the same as thresholding the the raw T at 3.2651, since the mapping from statistic values to P-values in the _vox_tstat is different at each voxel.
So, again, it's an apple and oranges problem... clusters found with a 3.5 threshold on a T image, vs. clusters found with a 0.001 on a nonparametric uncorrected P-value image. If you try -c 3.2651 you might see less of a difference between the two methods, but I would still expect to see some differences (Interpretation: No difference = Voxel-wise null distribution is homogeneous, T values mean the same thing everywhere; Big differences = Voxel-wise null distribution varies considerably across the image, T values actually have different significance in different regions).
Hope this helps.
-Tom
Thank you Gwenaelle and Tom!
Actually, the extent of the clusters at t=3.5 (i.e., "_tstat" files) is same to the extent of the clusters remaining significant after doing randomise -c 3.5 (i.e., "_maxc_tstat" files).
However, the extent of the clusters at p=0.001 (i.e., "_vox_tstat" files) is bigger the extent of the clusters remaining significant after doing randomise -c 3.5 (i.e., "_maxc_tstat" files).
Now, I guess the value in "_tstat" is not actually t-value in paired t-test.
Is it correct?
-----Original Message-----
From: Gwenaëlle DOUAUD [mailto:[log in to unmask]]
Sent: 2008年4月22日 12:52
To: [log in to unmask]Subject: Re: [FSL] FWE randomise and paired t-test
Hi Tom,
ahem, Steve pointed out some editing problems in my previous email.
So:
What Xiaochu did: Glm gui + Wizard for paired t-test in 50 subjects "before" and 50 corresponding subjects "after", so 51 EVs, DF 49.
What the problem is: when looking at the t-map in fslview, the extent of the clusters at t=3.5 is BIGGER than the extent of the clusters remaining significant after doing randomise -c 3.5.
Hope it makes sense this time!
Gwenaelle
--- En date de : Mar 22.4.08, Thomas Nichols <[log in to unmask]> a écrit :
> De: Thomas Nichols <[log in to unmask]>
> Objet: Re: [FSL] FWE randomise and paired t-test
> À: [log in to unmask]
> Date: Mardi 22 Avril 2008, 14h39
> Dear Xiaochu,
>
> I'm not sure I understand your question. Voxel and
> cluster inference will
> give different results, with cluster inference being better
> when the signal
> is 'clumpy', and voxel-wise being better for focal
> but intense signals.
> What works well on one dataset may not work well on another
> dataset.
>
> Hope this helps.
>
> -Tom
>
> On Sun, Apr 20, 2008 at 10:10 PM, Zhang, Xiaochu (NIH/NIDA)
> [F] <
> [log in to unmask]> wrote:
>
> > Hi FSL experts,
> >
> > I did some group t-test with randomise analysis. I
> usually used the
> > "Cluster-based thresholding corrected for
> multiple comparisons by using the
> > null distribution of the max (across the image)
> cluster size".
> >
> > It worked very well. However, recently I did a paired
> t-test analysis in
> > my new project. I made the parameter files with
> "Glm" wizard. However, the
> > multiple comparison correction made me confused. I
> found the t value (i.e.,
> > "-c" option) I set is not corrected any
> more. The activated clusters I got
> > looks very very small. For example, If we supposed
> t=3.5 (DF=48) means
> > p=0.001 in my study and I used "–c 3.5"
> during the analysis, the "maxc"
> > files show activated clusters are much smaller than I
> found in "vox" files
> > with p=0.001 threshold. I found similar results in a
> 2X2 ANOVA analysis,
> > too. Could you please do me a favor and give me some
> idea of it?
> >
> >
> >
> > Xiaochu Zhang Ph.D
> >
> > Visiting Research Fellow
> >
> > NIH/NIDA-IRP
> >
> > 251 Bayview Blvd
> >
> > Suite 200
> >
> > Baltimore MD 21224
> >
> >
> >
> > Tel: 443-740-2619
> >
> > Fax: 443-740-2734
> >
> >
> >
>
>
>
> --
> ____________________________________________
> Thomas Nichols, PhD
> Director, Modelling & Genetics
> GlaxoSmithKline Clinical Imaging Centre
>
> Senior Research Fellow
> Oxford University FMRIB Centre
__________________________________________________
Do You Yahoo!?
En finir avec le spam? Yahoo! Mail vous offre la meilleure protection possible contre les messages non sollicités
http://mail.yahoo.fr Yahoo! Mail