Hi

Please see http://www.fmrib.ox.ac.uk/fsl/melodic/index.html for an  
explanation: in the case of temp. concat melodic will not use the  
rank-1 approx during the iterations but will still (for various  
reasons)  plot the first Eigenvector (i.e. the result of a the rank-1  
approximation) in the html report as a summary time course. The web  
page above also will tell you how to get your hands on the individual  
(subject specific) time courses.

hth
christian


On 11 Jan 2008, at 22:12, Robert Kelly wrote:

> Hello,
>
> I have a question concerning a subject raised by Chris Bell in  
> November
> (see below), in particular concerning the following text
>
> "> This slightly
>> confuses me since with multi-session temporal concatenation mode,
>> I  thought
>> the time-course was assumed to be variable between subjects
>> (unlike  with
>> tensor ICA).
>
> that's right [answer from Christian Beckmann]"
>
> I ran resting state data for three subjects (170 volumes each) using  
> multi-
> session temporal concatenation (FSL 4.0) and I compared the output  
> to the
> output from the tensor ICA method without specifying a timeseries  
> model,
> finding the outputs (both spatial and temporal component patterns)  
> to be
> virtually identical.  This finding surprises me because I expect  
> that with
> tensor ICA each spatial component is produced in conjunction with  
> finding
> and fitting a common temporal course (the time course shown  
> graphically in
> the MELODIC output, 170 volumes long) for the three subjects, while  
> with
> temporal concatenation I expect that the time course involved is the
> concatenation of the time courses for each subject (3 x 170 volumes  
> long).
>
> Why would the outputs from the two approaches be the same?  If the
> temporal course used in the temporal concatenation approach (for a  
> given
> component) is the same as time course output shown for each  
> component (170
> volumes rather than 3 x 170 volumes long), then I don't think the  
> MELODIC
> temporal concatenation approach can be used for resting state data  
> or for
> any other data where a common stimulus temporal sequence is not  
> assumed to
> exist.
>
> Please help me understand this apparent contradiction.  I would  
> appreciate
> any helpful insights you could provide.  Thanks for your great  
> support.
>
> Best regards,
> Robert
>
>
> -----------------------------------------------------------------------
> Date:         Wed, 21 Nov 2007 16:41:20 +0000
> Reply-To:     FSL - FMRIB's Software Library <[log in to unmask]>
> Sender:       FSL - FMRIB's Software Library <[log in to unmask]>
> From:         "Christian F. Beckmann" <[log in to unmask]>
> Subject:      Re: group melodic analysis
> In-Reply-To:  <[log in to unmask]>
> Content-Type: text/plain; charset=US-ASCII; format=flowed; delsp=yes
>
> On 20 Nov 2007, at 20:34, Christopher Bell wrote:
>
>>
>>   Thanks much for your response. It was very helpful. What was it
>> that didn't make sense? Was it that I used contrasts in the concat
>> approach? I see now that comparing the effect sizes from a multi-
>> session decomposition may not be that meaningful, since the
>> contrasts are comparing effect sizes that have been based upon the
>> post-hoc largest eigen vector. Is that right?
>
> Not necessarily, in some cases it makes total sense
>>
>>   The contrasts used were A>B B>A A>0 B>0 . It is still a bit
>> unclear to me if these GLM contrasts are comparing differences in
>> the effect size (based on the largest eigen vector of the
>> timecourse) or differences in the spatial maps.
>
> These comparisons are never voxel-specific. They compare values in the
> subjetc-mode vector which is necessarily associated with the entire
> time course and the entire spatial map, i.e. they test if the areas
> shown in the spatial map (on average) show differences.
>
>> I am interested in the between group (subject domain) spatial map
>> differences. Would the Tensor ICA be more appropriate for this
>> purpose, even though it assumes a consistent timecourse?
>
> Not on resting data, unless you tansform the data into powerspectra
> first and are willing to assume that different RSNs have a clearly
> defined power-spectrum.
> hth
> christian
>
>>
>>
>> Chris Bell
>>
>> On Nov 19 2007, Christian F. Beckmann wrote:
>>
>>> Hi
>>>
>>> On 19 Nov 2007, at 22:38, Christopher Bell wrote:
>>>
>>>> I have run a group analysis on RSN data and am trying to correctly
>>>> interpret the output. I have specified 4 contrasts.
>>>
>>> I suspect you mean across the subject domain? This does not make
>>> that  much sense - the concat approach differs from full TICA in
>>> that the  rank-1 aroximation is not part of the estimation. It is,
>>> however,  being run post-hoc (after the components have been
>>> estimated free- form) because in some cases one might be interested
>>> in not using the  approximation during the estimation but still
>>> might be interested to  see how well the largest Eigen- time course
>>> represents variation  across subjects.
>>>
>>>> How are the subject mode
>>>> effect sizes calculated?
>>>
>>> These correspond to the factor loadings after taking all different
>>> time courses, assembling them into a matrix and calculating it's
>>> single largest EIgenvector.
>>>
>>>
>>>> Is their a technical report that discusses this?
>>>
>>> This rank-approximation is described in the T-ICA paper, see the
>>> technical report at
> http://www.fmrib.ox.ac.uk/analysis/techrep/tr04cb1/tr04cb1.pdf
>>>
>>>> I
>>>> am mainly interested in determining what is the cause of specific
>>>> subject
>>>> being an outlier for a component. I.e. is it a difference in this
>>>> specific
>>>> subject from the averaged spatial map or from the averaged time-
>>>> course?
>>>
>>>
>>>>
>>>>   Also, I believe the subject mode effect size has been referred
>>>> to as a
>>>> vector of spatial-temporal subject specific differences.
>>>
>>> Yes, in the context of tica
>>>
>>>> This slightly
>>>> confuses me since with multi-session temporal concatenation mode,
>>>> I  thought
>>>> the time-course was assumed to be variable between subjects
>>>> (unlike  with
>>>> tensor ICA).
>>>
>>> that's right
>>>
>>>> I don't understand the utility of an "averaged time-course" or
>>>> the contribution of a timecourse to a subject's effect size,
>>>> given  that the
>>>> subject's timecourse is not assumed to be consistent with other
>>>> subjects'
>>>> timecourses during the decomposition.
>>>
>>> You still might be interested in checking how consistent the effect
>>> is  in it's temporal characteristics across subjects, e.g. assume
>>> that you  do a learning task where (because people differn in the
>>> way they  learn) you do not want to use full tica to constrain the
>>> time course  to be the same. You therefore choose to estimate the
>>> components using  the concat apraoch but might still be interested
>>> to see how well the  'average' time course (rank-1 in fact, which
>>> is different) does  capture the full set of variation in subjects'
>>> learning.
>>>
>>>>
>>>> More broadly, if the t-tests turn up a difference between groups
>>>> is there
>>>> information about where this difference is spatially?
>>>
>>> In all cases (concat or tica) the estimated difference always is
>>> related to the entire spatial map associated with the time courses
>>>
>>>> Are these differences
>>>> calculated similarly to randomise?
>>>
>>> No, the report says that it's a simple GLm using ordinary leasts
>>> squares.
>>>
>>>> Is it possible to show the location of
>>>> between group differences in melodic
>>>
>>> It's related to the entire spatial map.
>>> hth
>>> christian
>>>
>>>> or is it best to run the groups
>>>> separatly and then compare the z-maps to find spatial-specific
>>>> differences
>>>> between groups? Sorry for so many questions!
>>>>
>>>> Chris Bell
>>>