Hi Anderson,
Im sorry, I did not explain in details the specificity of my design. It is a sparse sampling experiment (TR 10 sec, with a TA of 2.3 sec): We designed the experiment so that participants perform a task (listening, playing...) in a silent environment (while the scanner is not active during 7.7 sec), and we capture the bold response related to the task afterward (during the TA). Then only one volume is acquired per listening/playing/... trial, around 6 sec after the event of interest. Hence no convolution is needed in the Glm.
Also the parametric regressor correspond to the average error per trial (and not for each movement performed).
My question is why this parametric regressor (the averaged error) does not correlate with brain activity.
Another thing I should mention here is that I actually entered 2 parametric regressors (pitch error & time error) related to the playing regressor modeling the events: Should I orthogonalise both the playing non-parametric regressor and the parametric regressors (i.e. tick the "orthogonalise" button for all three regressors), or only the 2 parametric regressors with respect to the non-parametric one (i.e. tick the "orthogonalise" button for the two parametric regressors, as I did in the example I sent you) ? In that case, should I additionally orthogonalise the 2nd parametric regressor (e.g. time error) to the 1st one (e.g. pitch error) ?
Hope that's more understandable.
Thanks in advance
Indiana
________________________________
De : FSL - FMRIB's Software Library <[log in to unmask]> de la part de Anderson M. Winkler <[log in to unmask]>
Envoyé : lundi 9 mai 2016 05:55
À : [log in to unmask]
Objet : Re: [FSL] GLM analysis using behavioural measures
Hi Indiana,
I'm afraid I don't understand what is being coded. In a single run, in which 97 volumes are acquired, are the subjects asked to do the 4 different tasks, i.e., do they listen to the cello, then they play the cello with and without feedback, then imagine the cello? I see no convolution with the HRF, and it seems all the play events are being coded as single trials (?). It doesn't seem appropriate for a 1st level. I'd have thought that some key events during the playing, not the playing on its own right, that would be coded, and a comparison of the the conditions would be at a higher level.
All the best,
Anderson
On 6 May 2016 at 14:55, Indiana Wollman <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear Anderson,
Thanks for your prompt reply. Please find attached the 2 files (design.con & design.mat) of one participant 1st level (1 run).
Best,
Indiana
________________________________
De : FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> de la part de Anderson M. Winkler <[log in to unmask]<mailto:[log in to unmask]>>
Envoyé : vendredi 6 mai 2016 05:01:06
À : [log in to unmask]<mailto:[log in to unmask]>
Objet : Re: [FSL] GLM analysis using behavioural measures
Hi Indiana,
Could you send your design.mat and design.con (1st level, for a single, representative subject)? Thanks.
At any rate, no worries with the 2.3. If there is an issue, it isn't related to that.
All the best,
Anderson
On 5 May 2016 at 21:59, Indiana Wollman <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear Anderson,
Thank you for your reply.
As a reminder, I want to look at brain activity as a function of behavioural performance (ie. the relationship between how well subjects perform a task, and the resulting brain activation). Basically, I want to run parametric regression with FSL.
My experimental design is as follows:
- 4 Experimental runs, each containing 97 trials
- Within the 97 trials, there are 4 conditions
· 1: Listening to cello sequences
· 2: Playing the cello
· 3: Playing the cello without having auditory feedback
· 4: Imagining cello performance
For each of the Playing Trials (Conditions 2 & 3), the participants’ behavior is scored based on their pitch and tempo accuracy. For the remaining conditions, there is no behavioural score. As you suggested, I added a new EV file ("Play_behavior") in the full model setup of the first level that contains the behavioural scores on a per trial basis, in addition to the Playing text file containing 0’s and 1’s to model the events. I also tick the boxes "orthogonalise" in the full model setup, to orthogonalise the new EV containing the behavioral scores with respect to the EV modeling the events.
However, when I look at the results (third level in FSL), I have a nice image for the main contrast Play vs. rest, but no activation at all in the contrast Play_behavior vs. rest.
Do you know why? Is it due to the fact that the threshold (z = 2.3) I used is too restrictive for this regressor ?
Thank you in advance for your help,
Indiana
________________________________
De : FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> de la part de Anderson M. Winkler <[log in to unmask]<mailto:[log in to unmask]>>
Envoyé : mardi 22 mars 2016 05:04:28
À : [log in to unmask]<mailto:[log in to unmask]>
Objet : Re: [FSL] GLM analysis using behavioural measures
Hi Indiana,
Please, see below:
On 21 March 2016 at 19:21, SUBSCRIBE FSL Anonymous <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear list,
I am running a GLM analysis using FSL-FEAT 5.0.8, and I want to look at brain activity as a function of behavioural performance (ie. The relationship between how well subjects perform a task, and the resulting brain activation).
My experimental design is as follows:
- 4 Experimental runs, each containing 97 trials
- Within the 97 trials, there are 4 conditions:
· 1: Listening to cello sequences
· 2: Playing the cello
· 3: Playing the cello without having auditory feedback
· 4: Imagining cello performance
For each of the Playing Trials (Conditions 2 & 3), the participants’ behavior is scored based on their pitch and tempo accuracy. The score is currently stored as a floating value. The result of this, is that I end up with a continuous (dependent) variable for performance, which is not typically what you would use as a regressor. For the remaining conditions, there is no behavioural score.
My goal would be to use each individual behavioural score on a per-trial basis, as opposed to an average behavioural score on a per-subject basis. Once again, the end goal for this specific analysis is to observe the relationship between brain activation and task performance.
I have 3 main questions about how this should be implemented in FSL:
1. Should I add a new EV file in the full model setup of the first level called “Playing_Behaviour” that contains the behavioural scores on a per trial basis (in addition to the Playing text file containing 0’s and 1’s to model the events)
Yes. This qualifies as a parametric regressor: you'd enter one EV for the condition (playing) and one for the continuous score that indicates their pitch and accuracy.
OR should I simply replace the text file containing 0’s and 1’s with the Playing_Behaviour file.
No.
2. Does the data need to be demeaned
At the 1st level it isn't needed. Mean-centering is done automatically.
3. If it is the case that I enter an additional EV File every time I want to include behavioural information, can I enter as many EV Files as I want? Is there an upper limit to how many EVs you can include at the first level?
Yes, there is a theoretical limit, which is as many EVs as timepoints, but usually we don't want to be even close to that. Fewer is generally better, but with FMRI usually there are hundreds of timepoints so this is rarely an issue.
Hope this helps.
All the best,
Anderson
Thank you in advance for your help!
