Back to PLS HelpPosted on 12/05/08 08:09:25
Number of posts: 2
Hi,
I am wondering if anyone has any suggestion on how to deal with fMRI data that is not in sync with the experimental protocol? As I understand things it would be beneficial to have the fMRI sampling synchronized with the experimental protocol for PLS analysis, since the PLS software rounds down to the preceding scan index for any trial onset time. Any variability between sampling and trial onset would smear out the BOLD response when averaging across trials for a specific subject in a group analysis, right? Possibly with a “smear size” close to a full TR, since PLS always rounds down. When doing regression-based analyses the general advice is NOT to sync data sampling and experimental task, since variable sampling of the HRF gives you more info of the HRF shape. But now I am interested in doing a PLS analysis on non-synced data, but am unsure of how to best preserve the signal. Suggestions much appreciated!
Best regards
JohanReplies:
Posted on 12/07/08 09:31:58
Number of posts: 394
If I understand your question, I really don't see that there is a problem. Let me see if I understand your concern. For a given experiment, lets say you have a TR of 2 and you have events happening at 2 sec, 9 sec and 15 sec into a run. The events at 9 and 15 happen in the middle of a TR, so for PLS the onsets would be rounded down to volumes 4 and 7 (8 secs and 14 secs, respectively). Keep in mind that you are sampling every 2 seconds so you really have no data at the sub-TR level (unless you fiddle with slice timing, but thats tricky). The BOLD dynamics are such that deflections express themselves over 4-6 secs, at least from what I have seen published, so shifting the event onset back will not cause any appreciable damage. Now, if you reduce your TR to 1 second or less, then this "problem" goes away, but then there are other issues with SNR and field-of-view.
Its important to realize that this issue is not unique to PLS. All of the packages out there have to deal with this.
Another thing to consider, to put this into perspective, is that the same issue does arise in ERP/MEG studies. Neuroelectrical events occur at millisecond or submillisec ranges, but often data are sampled every 2 or 4 ms, so there is a bit of smearing of the recorded signal when the event of interest does not line up perfectly with the data acquisition.
Posted on 12/08/08 15:55:29
Number of posts: 2
Hi,
Thank you very much for your input! I believe that you have understood the problem and I agree with what you’re writing (though I suppose that variable sampling in a regression-based analysis can be seen as a feature rather than a bug, since the regression model can be made with infinite resolution and hence allow a perfect match between model and data, time-wise). It is interesting that you mention “fiddling with slice timing”, because that was the only thing I could think of as a possible “remedy”. Do you have any experience of this? Either way, if you are correct in that the problem isn’t much of a problem, I guess that trial-specific slice timing correction wouldn’t do much of a difference anyway…
Best
JohanLogin to reply to this topic.
