Hi Dan,

never tried running on an average subject so can't comment. All of  
these approaches imply different generative models, i.e. when using  
the average subject the underlying assumption is that every single  
data set for subject k, Y^(k), is represented via R common spatial  
and temporal modes independent of k: Y^(k)=\sum_{r=1}^R a_r \times  
b_r + \epsilon^(k)
That is, different data sets do not differ in structure, only in the  
stochastic noise. Such an approach is not able to model e.g. effects  
which only exist in one of the subjects (an image artefact, say) well.

When using a concatenation approach you'll be estimating one common  
and multiple different modes per component, e.g. when concatenating  
time series you effectively assume that spatial maps are identical  
but temporal dynamics associated with each spatial map might differ  
between subjects:
Y^(k)=\sum_{r=1}^R a^(k)_r \times b_r + \epsilon^(k)

note that now the time series a ends up being potentially different  
between subjects. If this is what you want to assume this might be  
the right model to use.
If you ran melodic separately you end up assuming that
Y^(k)=\sum_{r=1}^R a^(k)_r \times b^(k)_r + \epsilon^(k)
i.e. that every data set has it's own set of time courses and  
associated spatial maps - this then makes cross-subject comparisons  
quite difficult as you'd need some way of matching the spatial maps  
or time courses which always requires a bit of heuristics.

The advantages and disadvantages of a specific approach depend on  
what you want to assume about the underlying signals - as George Box  
has put it: "All models are wrong, some models are useful"

cheers
Christian


On 3 Oct 2006, at 14:07, Daniel Wolf wrote:

> Hi,
> While we await tensor-ICA, could you comment on the ads/disads of
> doing group ICA on concatenated (in time, across subjects) single
> subject studies, versus averaging across subjects in a group and doing
> ICA on an "average subject"?
> Thanks,
> Dan
>
> On 10/2/06, Christian Beckmann <[log in to unmask]> wrote:
>> Hi
>>
>> nope, the tensor-ICA approach is not part of FSL yet, it's currently
>> being implemented as c++ code and will be part of future versions.
>> The best way to proceed depends on what you're interested in. If you
>> want tpo perform a one- or two sample t-test between spatial maps you
>> could use a randomisation test (randomize on the command line)
>> hope this helps
>> best
>> christian
>>
>>
>>
>> On 2 Oct 2006, at 12:21, Paola Valsasina wrote:
>>
>> > Dear FSL experts,
>> >
>> > I am trying to run Melodic on resting state data using FSL version
>> > 3.2.
>> > I was able to run Melodic and produce IC maps on individual  
>> subjects.
>> > However, I am not sure on how to proceed for doing group analysis.
>> > My questions are:
>> > 1) Is there the possibility with FSL 3.2 (or 3.3) to perform a  
>> group
>> > analysis following the tensorial approach described in the
>> > Neuroimage paper?
>> > If yes, is there an user guide about the steps to be followed?
>> > 2) If not, once I obtain individual IC maps, which is the best
>> > method to
>> > select IC maps of interest and perform statistical inference on
>> > these maps
>> > (within a group or between two groups)?
>> > Best regards,
>> >
>> > Paola
>>
>> --
>>   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
>>

--
  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