Hi,
the order of compoennts after a standard decomposition is not
fundamentally meaningful - unlike a PCA decomposition the time
courses are not mutually orthogonal and therefore do not explain
separate amounts of variance.
Given that there is no natural order it's then sensible to order
components according to some criterion, e.g. in the current release
version melodic orders the components in order of decreasing amounts
of _uniquely_ explained variance. This then is meaningful...
Wrt number of components: yes, we believe that selecting the number
of components based on an estimate of the amount of non-Gaussianity
contained in the data is a good thing - picking a number at random
has various problems, some of which are summarised in the relevant
technical report.
Wrt group analysis: don't know what you mean by 'group analysis'. The
current release version does not do group ICA... for identifying
"corresponding" maps between decompositions. This is difficult but
can be achieved using e.g. spatial correlations (use avwcc for this).
The next release version will include multi-subject decompositions,
including Tensor-ICA, simple concatenation and PARAFAC, so the
problem of finding correspondences then is solved as part of the
decomposition... watch this space.
cheers
Christian
On 20 Dec 2006, at 20:08, Christina Hugenschmidt wrote:
> Hi All,
>
> I am interested in trying an ICA analysis of some of my data, and
> have been doing some reading. As a complete novice, though, I still
> have some questions. They are rather basic, but I am appreciative
> of any insights people could offer.
>
>
> Is the order of the components meaningful?
> In reading the technical papers about MELODIC, it seems that
> allowing FSL to choose the number of components is a good option.
> However, in reading the literature, it seems that people usually
> specify a set number of components. Which is the better option? If
> you specify a number of components, how do you choose a good
> number? My understanding from reading is that if a subject does not
> converge onto the specified number of components, that data is
> excluded from further group analysis. Is this so?
> Does anyone have recommendations for how to efficiently identify a
> desired network, so that that component can be used in a group
> analysis?
>
> Thank you!
>
>
> Christina
>
>
> Christina Hugenschmidt
>
> Graduate Student, Neuroscience Program
>
> Wake Forest University School of Medicine
>
> [log in to unmask]
>
> (336) 716-0972
>
>
>
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222551 Fax: +44(0)1865 222717
|