Hi,
On 27 Aug 2009, at 21:30, Lin Nga wrote:
> Hi,
>
> We've ran the same data set through Melodic's temporal concatenation
> several times and have gotten slightly different results each time.
> While I know there are some randomness to PICA, are these
> differences normal?
Yes - you should be worried if they're _exactly_ the same, because ICA
should be initialised with a different random mixing matrix each time
you run it.
> Each iteration were run with exactly the same settings, only varying
> by what were inputted for the post-stat matrix and contrast files.
> We had 47 subjects (inputted in the same order each time), forced
> the analyses to produce 20 components, turned off all options in the
> pre-stats (as these data have been denoised), and they were variance
> normalized. The results of the multiple iterations can be seen here: http://drop.io/melodic_iterations
>
> 1 - I accidentally mixed up the timeseries and the session/subject
> matrix and contrast files. Thus necessitating a second iteration.
> 2 - The matrix and contrast files are inputted correctly. The
> results were not the same as the first iteration so we ran more to
> see what was happening.
> 3 - Only the timeseries matrix and contrast files were entered.
> 4 - Only the group matrix and contrast files were entered.
> 5 - No matrix or contrast files were entered at all.
> 5.1 - A second iteration of 5
> 5.2 - No matrix or contrast files were entered at all and I started
> from a completely new melodic_gui/design.fsf (the other iterations,
> I have used the design.fsf from previous iterations but changing it
> in the gui).
>
> Questions:
>
> 1. The order of the components vary from one iteration to the next.
> Thus the component of interest's percent of explained variance
> varies by about 2%. Is this normal? If so, how would we decide which
> iteration to use?
The ordering of ICA components is not fundamentally controlled by the
ICA output - unlike with PCA, where it is natural to order according
to variance, there is no equivalent natural ordering of components -
hence you should not rely on (or be interested in) the specific
ordering of ICA components.
Having said that, that issue is separate from the fact that ICA will
give you (at least very slightly) different results each time you run
it, even if you re-order the components between two runs, to best
match the orderings.
The use of group matrix and contrast files has no effect at all on the
latter (the ICA results) - they are just used to help interpret the
ICA results.
>
> 2. In iteration 1 there were no GLM reports, as expected, but in the
> iterations where I did not enter any group GLM information
> (3,5,5.1,5.2) a GLM on subject/session-mode was reported. Is the PE
> here just the group mean and the copes (+/-) group mean contrasts?
> If not, where is Melodic getting this information?
Yes - this is just wrt the mean - though in the case of multi-subject
concat-ICA is not necessarily interpretable as this.
Cheers.
>
> Thank you!
> -Lin
---------------------------------------------------------------------------
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
---------------------------------------------------------------------------
|