I have a design with three conditions: positive, negative and neutral auditory stimuli lasting for 1 sec and presented as distractors in the delay period of a delay-matched to sample task. I am interested to study if the psychophysiological response to emo stimuli predict behavioural performance.
I have some doubt related to i) which basis function I should better use, ii) how to interpet the output file when I export the data and iii) what is the best choice for setting the duration of these events
i) I ran a first level analysis and I'm exploring the different basis functions scrf, scrf1, scrf2. But the data look somewhat different depending on which function I'm using, so I would need to make an informed choice here. On the basis of the type of design I outlined above, do you have suggestions as what is the best choice here?
ii) When I export the data for each subject (which I assume are beta values right?) I get extra columns when I use basis functions with derivatives.
Specifically, I set simple contrasts for each of these conditions
1) 1 0 0
2) 0 1 0
3) 0 0 1
I would expect 3 columns, instead I get 6 in the case of scrf1 and 9 in the case of scrf2.
I understand that the extra columns are each related to the time or time and dispersion derivatives.
But I'm not sure what is the order, i would assume that the first three columns are my beta values I should analyse right?
iii) I realise that the timing of my trial is not ideal for such a slow response as the SCR, but with your tool maybe it should be reasonably possible to model the response. I'm trying to model the data as I would do in an fMRI GLM analysis. My first thought was to just model from the vocalisation onset until the end of the trial but then I read on the manual:
"[...] the duration field can be omitted if you assume zero duration (which is normally the case)."
I don't understand why would would the event duration assumed to be zero?
Thanks in advance for your help and sorry if some questions are silly but better to ask them than making huge mistakes.
Best
Antoinette
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
(i) According to Bach, Friston, Dolan, Biological Psychology 2013, scrf1 together with the reconstruction option provides the best sensitivity in a GLM to distinguish negative/neutral or positive/neutral iamges, consistently across 2 experiments. So this seems to be the best choice.
(ii) The order of the regressors is (1) canonical function condition 1 (2) time derivative (3) dispersion derivative (4) canonical function condition 2, etc. but your best choice is to use the reconstruct option and then you should get only 3 columns, one for each condition
(iii) Although your events have a non-zero duration, sudomotor nerve bursts are typically compact and short. We assume a zero-duration event which causes a compact sudomotor burst. If you model events with a non-zero duration, the assumed sudomotor burst also becomes longer - but this is probably not the case physiologically. It is more likely that you get short bursts during the event, they could be varying in latency, or even repeated bursts. Non-linear models (DCM) would be ideal to explore this possibility, but with an event duration of 1.5 and 2 s, a GLM may even suffice. However, you need to model all events, not just the ones you are interested in.
Does that make sense at all?
Best wishes and a happy new year
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If I use the GLM, I should model the onset of all events but with zero duration? So a regressor for memorandum and one for the probe, all of them with zero duration. Right?
I might give it a try with the DCM but first I want to get this straight and then I'll start reading about non-linear models.
Thanks so much for your support!
Best
Antoinette
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear Dominik,
I have a design with three conditions: positive, negative and neutral auditory stimuli lasting for 1 sec and presented as distractors in the delay period of a delay-matched to sample task. I am interested to study if the psychophysiological response to emo stimuli predict behavioural performance.
Trial Timing:
Memorandum(2 sec) - ISI(0.5 sec) - emotional auditory stimuli (1 sec) - ISI(0.5 sec) - Probe (1.5) - ITI (3sec)
I have some doubt related to i) which basis function I should better use, ii) how to interpet the output file when I export the data and iii) what is the best choice for setting the duration of these events
i) I ran a first level analysis and I'm exploring the different basis functions scrf, scrf1, scrf2. But the data look somewhat different depending on which function I'm using, so I would need to make an informed choice here. On the basis of the type of design I outlined above, do you have suggestions as what is the best choice here?
ii) When I export the data for each subject (which I assume are beta values right?) I get extra columns when I use basis functions with derivatives.
Specifically, I set simple contrasts for each of these conditions
1) 1 0 0
2) 0 1 0
3) 0 0 1
I would expect 3 columns, instead I get 6 in the case of scrf1 and 9 in the case of scrf2.
I understand that the extra columns are each related to the time or time and dispersion derivatives.
But I'm not sure what is the order, i would assume that the first three columns are my beta values I should analyse right?
iii) I realise that the timing of my trial is not ideal for such a slow response as the SCR, but with your tool maybe it should be reasonably possible to model the response. I'm trying to model the data as I would do in an fMRI GLM analysis. My first thought was to just model from the vocalisation onset until the end of the trial but then I read on the manual:
"[...] the duration field can be omitted if you assume zero duration (which is normally the case)."
I don't understand why would would the event duration assumed to be zero?
Thanks in advance for your help and sorry if some questions are silly but better to ask them than making huge mistakes.
Best
Antoinette
Hi Antoinette
most happy to help.
(i) According to Bach, Friston, Dolan, Biological Psychology 2013, scrf1 together with the reconstruction option provides the best sensitivity in a GLM to distinguish negative/neutral or positive/neutral iamges, consistently across 2 experiments. So this seems to be the best choice.
(ii) The order of the regressors is (1) canonical function condition 1 (2) time derivative (3) dispersion derivative (4) canonical function condition 2, etc. but your best choice is to use the reconstruct option and then you should get only 3 columns, one for each condition
(iii) Although your events have a non-zero duration, sudomotor nerve bursts are typically compact and short. We assume a zero-duration event which causes a compact sudomotor burst. If you model events with a non-zero duration, the assumed sudomotor burst also becomes longer - but this is probably not the case physiologically. It is more likely that you get short bursts during the event, they could be varying in latency, or even repeated bursts. Non-linear models (DCM) would be ideal to explore this possibility, but with an event duration of 1.5 and 2 s, a GLM may even suffice. However, you need to model all events, not just the ones you are interested in.
Does that make sense at all?
Best wishes and a happy new year
Dominik
Dear Dominik,
All the best for the new year indeed!!
Thanks so much for your detailed answer. Just a few clarifications, just to be sure:
ii) Where do I find the reconstruct option?
iii) So if I understand correctly, given my design:
Trial Timing:
Memorandum(2 sec) - ISI(0.5 sec) - emotional auditory stimuli (1 sec) - ISI(0.5 sec) - Probe (1.5) - ITI (3sec)
If I use the GLM, I should model the onset of all events but with zero duration? So a regressor for memorandum and one for the probe, all of them with zero duration. Right?
I might give it a try with the DCM but first I want to get this straight and then I'll start reading about non-linear models.
Thanks so much for your support!
Best
Antoinette
You can find the 'reconstruct' option in the constrast module.
You could also download the development version using an SVN client; it has a much nicer GUI and will be released very soon.
Best
Dom