Hi,
On 5 Jun 2007, at 17:40, Antonios - Constantine wrote:
> Dear fsl users,
>
> I have 8 control groups and 14 Parkinson disease subjects and after
> using
> siena and sienax i wanted to use the randomise tool in order to
> localize
> where there's a difference in atrophy in these groups..
> I've read both the randomize manual (including the example in the
> practical
> instructions) and the Holmes &Nichols paper for nonparametric
> permutation
> test, though i still have a lot of questions about the
> “translation” of the
> results after using the randomise tool. I would really appreciate
> your help
> once more...
> My questions are the following:
>
> 1)Is it a problem if the number of controls is not the same with the
> patients number? if yes is it recommended to reduce the number of
> patients to 8?
No, it's fine for the numbers to be different.
> 2)The outputs of randomise are 16 different statistics. 4
> sienar_tstat, 4
> sienar_maxc_tstat, 4sienar_max_tstat and 4 sienar_vox_tstats. If i
> understood well, first we check the p-values (which are filtered
> through
> fslview in the range [0.949,1] in order to see p-values <0.05) from
> sienar_maxc_tstat1 and sienar_maxc_tstat2 and we detect if there's
> a group
> difference. In your example there are no clusters with voxels in
> sienar_maxc_tstat2 with p-value <0.05 while there are some clusters in
> sienar_maxc_tstat1 with p<0.05..So this means that the control-
> patient group
> shows a positive value?
Yes, within those voxels.
> What would be the conclusions if there were also
> clusters in sienar_maxc_tstat2 with p<0.05?
patient>controls in _those_ voxels (they will not be overlapping with
the first set above!)
> What about the
> sienar_maxc_tstat3 and sienar_maxc_tstat4?what kind of info can we
> extract
> from them?
It depends what contrasts you selected for tstat3 and 4. If they are
the two group means they are just asking - where is the group mean
different from zero.....probably not interesting in this case.
> 3)The next step is to check the sienar_tstat3 and sienar_tstat4 and
> disambiguate what's going on where there's a group difference.Are
> these
> stats corrected for multiple comparisons?
Yes - see the randomise manual http://fsl.fmrib.ox.ac.uk/fsl/
randomise/ - that's what maxc and max are.
> According to your example these
> stats gives us the info about what each group is doing
> separately...when i
> load them in fslview independently i see 2 different colors for
> sienar_tstat3(red yellow filtered with fslview in a range of
> [0.5,3].why do
> we need to filter the output in this range?and why do we need two
> colors to
> depict them in fslview?these intensities are definitely not p-
> values but
> what exactly are?)
You can choose whatever colourmap you like. If you choose "Red" then
you will just see different brightnesses of red.
If you click on the (i) and select Red and then turn on the second
(negative LUT) just below that and select Blue, then set the
intensity display range to say 2:5 you will see negative values <-2
in Blue and positive values >2 in Red.
The tstat3 image is the raw t-statistic before randomisation is
actually run to get conversion to p-values.
> and intensities with bigger magnitude in yellow and lower
> magnitude in red...Exactly the same in the sienar_tstat4 with
> colors blue
> and light blue...How can i “explain” these results ? What kind of
> information does these sienar_tstats gives us? what about the
> sienar_tstat1
> and sienar_tstat2?what kind of info can we have from them?
tstat1 (group difference) shows you where one group has atrophy
values that are greater than the other - to interpret fully you also
need to look at 3 and 4 to find out whether each is actually positive
or negative (you can't tell that just from the difference).
> 4)The cluster threshold in your example is set to 1...why did you
> choose
> this threshold?
Because in that example the effect was weak. setting the cluster-
forming threshold is arbitrary and I'm afraid there's no good answer
to how to set it - many people just set it at say 2.3 or 3 and leave
it at that. Probably a good idea to search the email list archives
for more on this - it's a big subject!
> what's the range of the values that this threshold can
> take?Is it a low threshold or a large one?do we have to use more
> than one
> cluster thresholds each time? I read that if we choose a low one we
> loose
> intense focal signals, while if we choose a high one then we loose low
> intensity signals..
> 5)What about the sienar_max_tstats?what kind of information can we
> extract
> from these stats?we use them only to localize the voxels of
> significance
> since this is the weakness of a suprathreshold tests?
Again - see the manual - these are voxelwise p-values corrected for
multiple comparisons - if you get results there then fine - but most
likely this won't show up anything significant, which is why we also
use the cluster-based stats.
> As for the
> sienar_vox_tstats how useful these data can be since there're
> uncorrected
> for multiple comparisons and what kind of info we extract from them?
If you knew exactly where you were interested in looking then you
wouldn't need to do multiple comparison correction - so you could
then use these p-values.
Alternatively, you can feed these uncorrected p-values into FDR for
an alternative method of multiple comparison correction.
> thanks a lot in advance and i'm really sorry if i tired you with this
> extended e-mail but i couldn't find some information to help me on my
> previous questions.
No problem - cheers, Steve.
>
> Antonios-Constantine Thanellas
------------------------------------------------------------------------
---
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
------------------------------------------------------------------------
---
|