One final question about TBSS syntax--
In what scenarios should I use the -c <thresh> versus the -F <thresh> in
order to control for multiple comparisons. According to the very
informative response I got from Dr. Smith (see below), the model I set up
had some "t" type contrasts and some that appeared to be properly controlled
for using an F based statistic. Would I just run TBSS twice or use both
options on the command line and just look at the corresponding output files?
---------------------------------------------------------------------
Sure - these are just simple contrasts that mean average_drug_user -
controls etc. and don't need f-tests.
For an f-test across 'all' comparisons you probably just need an f
across the top two contrasts, i.e
F
1 -1 0 x
1 0 -1 x
To use this option, use -c <thresh> for t contrasts and -F <thresh>
> for F contrasts, where the threshold is used to form supra-threshold
> clusters of voxels.
----------------------------------------------------------------------------------------------------------
On Fri, Mar 28, 2008 at 3:51 AM, Steve Smith <[log in to unmask]> wrote:
> Hi,
>
> On 26 Mar 2008, at 21:24, David Gutman wrote:
> > Greetings fellow FSL users,
> >
> > I've recently been looking over an old thread about 3 group TBSS
> > Comparisons and had some questions about the command syntax. I am
> > doing a 3 group comparison, 2 groups are drug dependent (with the
> > differentiating factor being whether they relapsed during the 6
> > months to 1 year followup of our study), and the other group being
> > controls.
> >
> > It seemed like in the original thread from ~ a month ago, in a
> > message from Tim in response to a question by David K, it was
> > recommended to set up the following contrasts
> >
> > 1 -1 0
> > 1 0 -1
> > 0 1 -1
> >
> > It seems like from my (comparatively limited) experience with TBSS,
> > it appears the statistical output in the tbss_max_vox_stats* images
> > are only one sided-- i.e. indicate voxels where A>B and so to
> > actually do all the comparisons you would also need to add
> >
> > -1 1 0
> > -1 0 1
> > 0 -1 1
> >
> > to the design.mat file you set up in the GLM. Is that
> > correct..........?
>
> Correct.
>
> > Also I am a bit confused about the exact syntax to use for randomise
> > and fear I used the wrong commands---- perhaps I am just confusing
> > the nomenclature. I initially ran my 3 group comparison using
> >
> > randomise -i all_FA_skeletonised.nii.gz -o tbss -m
> > mean_FA_skeleton_mask.nii.gz -d design_3groups.mat -t
> > design_3groups.con -c 3
> >
> > I recently realized I need to add -f design_3groups.fts to the
> > randomise command as well, but now that I think about it more, I am
> > not sure if the "-c 3" is correct. Since there are 3 groups (i.e.
> > F tests..?) do I actually want to use the -F option .... and if so
> > what threshold would I use? In the contrats above I am just going t-
> > tests between groups, but in my actual design matrix I am also
> > intrasted in contrasts that compare between more than 2 groups (i.e.
> > comparing simply ALL_DRUG_USERS - CONTROLS)
> >
> > i.e. something like ..
> > 0.5 0.5 -1
> > -0.5 -0.5 1
>
> Sure - these are just simple contrasts that mean average_drug_user -
> controls etc. and don't need f-tests.
>
> For an f-test across 'all' comparisons you probably just need an f
> across the top two contrasts, i.e
>
> F
> 1 -1 0 x
> 1 0 -1 x
>
> (you don't need the third contrast in this as it is a linear
> combination of the first two)
>
> Wrt choosing a threshold for the F option: as always, this is
> arbitrary, you'll need to see what works reasonable by looking at the
> raw fstat image.
>
>
> >
> >
> > --------- FROM THE FSL WEB PAGE
> > To use this option, use -c <thresh> for t contrasts and -F <thresh>
> > for F contrasts, where the threshold is used to form supra-threshold
> > clusters of voxels.
> > ----------------------------------------------------
> >
> >
> > Could anyone provide some clarification?
> >
> >
> > FINALLY--- if I want to control for multiple comparisons, but want
> > to exclude part of the brain a priori (there's a bad scanner
> > artifact in some of my subjects in the back of the brain related to
> > motion/scanner fibrations), so I just want to completely ignore a
> > chunk of the occipital lobe. Can I just mask that out of my -m
> > mean_FA_skeleton_mask.nii.gz, and so have it only correct for
> > multiple comparisons across the regions of the brain that don't have
> > obvious confounds like that.
>
> Yes, that's fine.
>
> Cheers.
>
>
>
>
> >
> >
> > DG
> >
> >
> > Thanks in advance,
> >
> > DG
> >
> >
> > --
> > David A Gutman, M.D. Ph.D.
> > Department of Psychiatry & Behavioral Sciences
> > Emory University School of Medicine
> >
> > On Fri, Mar 7, 2008 at 12:08 PM, Thomas Nichols <[log in to unmask]
> > > wrote:
> > David,
> >
> > I have two more questions re: -f design.fts.
> > 1) Do I need to include both the -f design.fts and -t design.con,
> > then, or just one? I would think both, but I just want to be sure.
> > Again, I want the t-test and f-test.
> >
> > Yes, you need both, since the F tests are defined in terms of the t-
> > tests. Ideally, if you didn't care about the t-tests, we'd have a
> > way to have randomise to skip the t-tests, but we don't have such an
> > option.
> >
> > 2) Can you tell me what one gains by including the f-tests if you
> > already including the t-tests. To my understanding, the F-test will
> > only tell you that there are differences between groups, but not
> > where these differences are (that is the purpose of the t-test). So,
> > by adding the -f design.fts command, what benefit will this give a
> > person for results interpretation?
> >
> > F tests give a single answer to the 'Any Difference' question, while
> > the t-tests give 6 answers (A<B,B<A,A<C,A>C,B<C,B<C, even though
> > there's only really 4 linearly independent answers). Those multiple
> > t-tests give multiple opportunities for false positives to occur,
> > and require a Bonferroni or some other correction.
> >
> > Hence, if 'Any Difference' is the primary question, then the F is
> > the standard answer. But, if you *did* discover any differences,
> > you could use the t-tests to characterize the differences in a post
> > hoc fashion (w/out formal control of false positives).
> >
> > -Tom
> >
> > ____________________________________________
> > Thomas Nichols, PhD
> > Director, Modelling & Genetics
> > GlaxoSmithKline Clinical Imaging Centre
> >
> > Senior Research Fellow
> > Oxford University FMRIB Centre
> >
> >
> >
>
>
>
> ---------------------------------------------------------------------------
> 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<http://www.fmrib.ox.ac.uk/%7Esteve>
>
> ---------------------------------------------------------------------------
>
--
David A Gutman, M.D. Ph.D.
Department of Psychiatry & Behavioral Sciences
Emory University School of Medicine
