I'm running a DCM on my fear conditioning study and wanted to double check something.
I've got 3 CSs - a CS+ that is always followed by a shock 12s later, a CS- which is never followed by a shock, and a CST which is always followed by a shock but either 6, 8 ,10, or 12s later. My ISIs are also variable - 4, 6, 8 or 10s.
When setting up my model I followed the DCM tutorial in that I made a timing file consisting of the onset & offset of all the CSs where the offset was when the shock occurred (or 12s after the CS- for no shock) in event{1,1} and the onset of all fixed shock/noshock times in event{1,2} (without offsets).
Conceptually I understand that setting the onset/offset tells the model the response to the CS will occur in this window. My concern is that since my ISIs can be very-short relatively-speaking I'm worried the shock 'onset' may affect the measurements of the peak response to the next trial. So for example if trial 1 is a CS+ and trial 2 is a CS- and the ISI between them is 4s, I have seen the response to shock begin at 3s post-shock-onset with the CS- being presented at 4s post-shock-onset so there will obviously be overlap. And if I am spec ifying my response window to begin at CS onset and the maximum response in the window is due to shock carry-over, will this be reported in error or does the modeling account for this given I have specified shock onset time in event{1,2}.
I am curious about this because I seem to have some very odd peak latency values in my paramater output (i.e. .004, .05, basically a handful of values below 1s) and I am curious as to if this is due to shock contamination or if my data is just noisy as there "only" 13/72 trials for which the latency is below 1s but I am used to seeing peak latency values around 6s post-CS-onset and even 13 seems like kinda a lot.
Follow-up: If this IS due to shock contamination can I account for it in any way with PsPM?
Thank you Dominik and others. You guys are the best.
Lauren
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
if you model two responses that are 4 s from each other, then as you say the first one will overlap with the second. In fact, the peak of the first response will occur during a time when the second one is initiated.
Note, however, that this does not mean they are not at all separable. The second response will only show up in SCR 1.5-2 s after the second event, when the first one is already decaying. If you simulate data with two eSCR, you will see that if your subject's response perfectly matches the SCRF, and there is no measurement noise, then the two response amplitudes are highly separable. Of course, these assumptions will not be met, and so there can be some degree of degeneracy between the two responses, i.e. different amplitude parameter combinations can yield the same data fit; but I don't think with a 4 s ITI this will be too problematic (see Gerster et al. 2018 Psychophysiology, a study with direct SN nerve stimulation, for the upper bound on stimulation frequency).
The more crucial aspect here is the dispersion of the anticipatory SN burst. PsPM by default sets the maximum dispersion (SD of a Gaussian impulse) to 1/2 of the anticipation window. With a 0 s peak latency and a 6 s SD, the Gaussian impulse and the ensuing SCR will overlap with the previous response, and this could be problematic. In fact, although the initial paper (Bach, Daunizeau et al. 2010) included a 16 s anticipation window, the sample size was rather small, and the conclusion that the model fits different anticipation windows equally well should be taken with a grain of salt. We are currenlty improving models for long anticipation windows.
Pragmatically, I'd suggest checking the dispersion value on the trials with short latency. If it is rather long, then you may have a problem. I would suggest reducing the upper bound on SN burst dispersion, either by fixing it to a physiologically plausible value (a new option available on the code repository) or by hacking into pspm_dcm_inv, line 591:
Hi Dominik:
I'm running a DCM on my fear conditioning study and wanted to double check something.
I've got 3 CSs - a CS+ that is always followed by a shock 12s later, a CS- which is never followed by a shock, and a CST which is always followed by a shock but either 6, 8 ,10, or 12s later. My ISIs are also variable - 4, 6, 8 or 10s.
When setting up my model I followed the DCM tutorial in that I made a timing file consisting of the onset & offset of all the CSs where the offset was when the shock occurred (or 12s after the CS- for no shock) in event{1,1} and the onset of all fixed shock/noshock times in event{1,2} (without offsets).
Conceptually I understand that setting the onset/offset tells the model the response to the CS will occur in this window. My concern is that since my ISIs can be very-short relatively-speaking I'm worried the shock 'onset' may affect the measurements of the peak response to the next trial. So for example if trial 1 is a CS+ and trial 2 is a CS- and the ISI between them is 4s, I have seen the response to shock begin at 3s post-shock-onset with the CS- being presented at 4s post-shock-onset so there will obviously be overlap. And if I am spec ifying my response window to begin at CS onset and the maximum response in the window is due to shock carry-over, will this be reported in error or does the modeling account for this given I have specified shock onset time in event{1,2}.
I am curious about this because I seem to have some very odd peak latency values in my paramater output (i.e. .004, .05, basically a handful of values below 1s) and I am curious as to if this is due to shock contamination or if my data is just noisy as there "only" 13/72 trials for which the latency is below 1s but I am used to seeing peak latency values around 6s post-CS-onset and even 13 seems like kinda a lot.
Follow-up: If this IS due to shock contamination can I account for it in any way with PsPM?
Thank you Dominik and others. You guys are the best.
Lauren
Hi Lauren
if you model two responses that are 4 s from each other, then as you say the first one will overlap with the second. In fact, the peak of the first response will occur during a time when the second one is initiated.
Note, however, that this does not mean they are not at all separable. The second response will only show up in SCR 1.5-2 s after the second event, when the first one is already decaying. If you simulate data with two eSCR, you will see that if your subject's response perfectly matches the SCRF, and there is no measurement noise, then the two response amplitudes are highly separable. Of course, these assumptions will not be met, and so there can be some degree of degeneracy between the two responses, i.e. different amplitude parameter combinations can yield the same data fit; but I don't think with a 4 s ITI this will be too problematic (see Gerster et al. 2018 Psychophysiology, a study with direct SN nerve stimulation, for the upper bound on stimulation frequency).
The more crucial aspect here is the dispersion of the anticipatory SN burst. PsPM by default sets the maximum dispersion (SD of a Gaussian impulse) to 1/2 of the anticipation window. With a 0 s peak latency and a 6 s SD, the Gaussian impulse and the ensuing SCR will overlap with the previous response, and this could be problematic. In fact, although the initial paper (Bach, Daunizeau et al. 2010) included a 16 s anticipation window, the sample size was rather small, and the conclusion that the model fits different anticipation windows equally well should be taken with a grain of salt. We are currenlty improving models for long anticipation windows.
Pragmatically, I'd suggest checking the dispersion value on the trials with short latency. If it is rather long, then you may have a problem. I would suggest reducing the upper bound on SN burst dispersion, either by fixing it to a physiologically plausible value (a new option available on the code repository) or by hacking into pspm_dcm_inv, line 591:
repmat(foo(:)/2, 1, size(u, 2)) - settings.dcm{1}.sigma_offset
where the 2 should be replaced with a larger number for stronger constraints. (foo is the duration of the anticipation window).
Hope this helps
Dominik