Hi - yes the next release of FSL will contain f-tests in randomise.
Cheers.
On 30 Apr 2007, at 12:36, Ged wrote:
> Hi Torsten,
>
> Sorry about my rubbish code! I was completely wrong about only one
> of the
> images being non-zero at each voxel, so it isn't correct to add the
> images
> together. Instead, I think taking the most significant p-value
> (maximum of
> the two 1-p images) and then doubling this is better. A slight
> complication
> is that there can be one-sided p-values over 0.5 (up to 1), which
> for a
> two-sided test, end up as being over 1. The attached script should do
> roughly the right thing...
>
> However, I'm not sure that combining two one-tailed permutation
> tests into a
> two-tailed one like this is strictly correct. I think randomise
> would have
> to be changed to look at the absolute value in comparison to the
> permutation
> distribution of the (max, if corrected) absolute value. But the
> differences
> would probably (I hope) be quite minor.
>
> For your other question: treating any t-contrast as an F-contrast
> gives a
> two-sided test, because the F-statistic is just the square of the
> t-statistic (hence positive or negative t go to positive F) in this
> situation (i.e. the hypothesis degrees of freedom = 1). However, F-
> contrasts
> are not (yet) available in randomise -- see:
>
> http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0702&L=FSL&P=R3728
>
> FEAT or SPM could use F-contrasts to give you a two-tailed test,
> however,
> this would use parametric stats again. As Jesper and I mentioned
> elsewhere
> in this thread, there may be advantages to non-parametric
> permutation-based
> testing as used in randomise (or SnPM) depending on your data (e.g.
> small
> subject groups, and/or not very normally-distributed errors), so
> you may
> want to persevere with randomise (or wait for the next version?!
> Steve --
> any comments on the likely time-scale?)
>
> Best,
> Ged.
>
>
> Torsten Ruest wrote:
>> Dear Ged,
>>
>> thanks very much for your suggestions. I tried the calculations, but
>> unfortunately, it doesn't work out :( I hope I don't bother you too
>> much, but if you could help me with this, I would be very glad!
>>
>> for a given voxel with the initial intensity in the max_tstat_1 of
>> 0.980519, and max_tstat_2 of 0
>>
>> avwmaths ${fnm}1 -mul -1 -add 1 -mul 2 ${fnm}1_p2
>> avwmaths ${fnm}2 -mul -1 -add 1 -mul 2 ${fnm}2_p2
>>
>> gives me max_tstat_1_p2 value of 0.0389611
>> max_tstat_2_p2 value of 2
>>
>> avwmaths ${fnm}1_p2 -add ${fnm}2_p2 ${fnm}_p2
>>
>> added together, I get for max_tstat_p2 a value of 2.0389611
>>
>> avwmaths ${fnm}_p2 -mul -1 -add 1 ${fnm}
>>
>> this gives me then an intensity of -1.03896
>>
>> Do you know of a solution for this problem?
>>
>>
>> Alternatively, I saw in the SPM mailing list messages from you, where
>> you say that including a f-test for either of the 2 contrasts, (1
>> -1) or
>> (-1 1), would be equivalent to a 2-tailed t-test. I looked through
>> the
>> fsl glm setup, and basically I could include a f-test too. Well it
>> didn't work, either I didn't set it up right, or fsl is behaving
>> differently than SPM.
>>
>> But luckily enough, I would have access to SPM, and I will check out
>> this possibilty on Monday. I am not sure if that will work,
>> because my
>> data is animal data, and I am currently not sure if there is just
>> a .mat
>> file created (and projected on a human glas brain), or actual stats
>> images, as in fsl, which I could overlay with another software.
>>
>> Anyway, enjoy your weekend. I really appreciate your help.
>>
>> Cheers,
>>
>> Torsten
>>
>>
>
> <to-two-tailed>
------------------------------------------------------------------------
---
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
------------------------------------------------------------------------
---
|