Print

Print
Hi Christian,

Thanks for the suggestion to use temporal concatenation -- it makes sense given the structure of the data. I suppose my confusion is how to specify the proper timeseries and subject/session design matrices for melodic. More specifically, because there are 4 sessions for each subject and each session has a different time series design matrix (although the same stimulus types), a single time series file doesn't describe the runs. In traditional event-related analyses (e.g., using FEAT), I would tend to follow the multi-session, multi-subject approach where lower-level stats are estimated for each run and are then combined per subject at the second level. I'm not sure how to achieve this sort of thing using the temporal concatenation ICA approach. More specifically, I want to know about how extracted ICA components relate to particular stimulus types (e.g., reward cue, loss cue in a gambling task) in the paradigm, but this requires a time series design matrix.

What would you suggest to resolve this issue?

Thanks very much,
Michael

On Tue, Jun 29, 2010 at 5:08 PM, Christian F. Beckmann <[log in to unmask]> wrote:
Hi

If the timecourses associated with any effect are not identical between subects/sessions then TICA is not appropriate. You can use the 'temporal concatenation' approach instead
hth
Christian



On 28 Jun 2010, at 21:07, Michael Hallquist wrote:

> Dear FSL users,
>
> I am learning melodic in order to undertake an ICA of an event-related task using an existing dataset. Although I have read several papers on this topic and have explored the melodic practical from the FSL course, I am uncertain whether Tensor ICA or Temporal Concatenation ICA is appropriate for my data.
>
> I have approximately 30 subjects, 15 adults and 15 adolescents who completed four runs of the same task. The task was pseudorandomized across runs such that all participants saw exactly the same four runs, but the order of stimulus presentation differed per run. Because all participants received the same stimulation, I had hoped to use Tensor ICA. But because sessions and subjects are both higher-level effects, I didn't know whether this could be done. One alternative I had considered is temporally concatenating each participant's data and running a multi-subject, single-session (from the perspective of melodic) Tensor ICA.
>
> What would you advise?
>
> Thanks very much for your help,
> Michael