I've been trying to use the DCM non-linear model to look at a fear conditioning data set that I have. I've set up a batch model based on the tutorial (in the manual) and copying code from the view code function in the batch editor.
Specifically, I am looking at an acquisition sequence with CS+US-, CS-US-, CS+US+. The electric shocks come at the end of the 3.5 colored-square CS, the ITI's are 10-12 seconds. There are 22 trials over the sequence.
The first level model reliably gives me a massive spike in the flexible response amplitude for the last trial (trial #22) across all of the subjects (think 1-2 orders of magnitude), regardless of whether that trial happens to be a CS+US- or a CS-US- for that subject.
At first, I thought that it could capturing response from the subject seeing a "You have finished screen", but that screen is well outside of the window that I specified in the event-timing files. Further, I re-ran the model with the time period of that screen trimmed it out of the SCR data.
I cannot figure out why I am seeing this, and I was wondering whether you had any advice here, and could help me out. I can provide any information as needed (code, output, clarification, etc.).
Thanks again,
Best,
Eric
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
it would be great if you could send a DCM file for us to have a look at. This almost sounds like a data or preprocessing issue, and the DCM file will contain all the relevant diagnostic information for us.
Thanks and best wishes
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
my first guess is that this happens because there is not much data left after the last event (3 s). Specifying a large response amplitude has comparatively little impact on the model fit.
However, this behaviour should still not happen, and I'll investigate whether it can be avoided. As a quick fix, if you happen to have more data at the end, then don't cut it away.
Best
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I found the solution to the problem. There were some conditions under which the priors & starting values for the estimation of the amplitude could be inappropriately large. In the absence of meaningful data (as in your last trial), the estimation will fall back to this prior.
I have fixed this bug, and the amplitude estimates for the last trial will be in the same range as the others. Note however that they will still not be interpretable and just reflect the amplitude from the response averaged across all trials, because there is not enough data to estimate the amplitude properly.
To get access to the bugfix, update from the code repository.
Hope this helps
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello Dr. Bach,
I've been trying to use the DCM non-linear model to look at a fear conditioning data set that I have. I've set up a batch model based on the tutorial (in the manual) and copying code from the view code function in the batch editor.
Specifically, I am looking at an acquisition sequence with CS+US-, CS-US-, CS+US+. The electric shocks come at the end of the 3.5 colored-square CS, the ITI's are 10-12 seconds. There are 22 trials over the sequence.
The first level model reliably gives me a massive spike in the flexible response amplitude for the last trial (trial #22) across all of the subjects (think 1-2 orders of magnitude), regardless of whether that trial happens to be a CS+US- or a CS-US- for that subject.
At first, I thought that it could capturing response from the subject seeing a "You have finished screen", but that screen is well outside of the window that I specified in the event-timing files. Further, I re-ran the model with the time period of that screen trimmed it out of the SCR data.
I cannot figure out why I am seeing this, and I was wondering whether you had any advice here, and could help me out. I can provide any information as needed (code, output, clarification, etc.).
Thanks again,
Best,
Eric
Hi Eric
it would be great if you could send a DCM file for us to have a look at. This almost sounds like a data or preprocessing issue, and the DCM file will contain all the relevant diagnostic information for us.
Thanks and best wishes
Dominik
Okay, I've attached an example first-level model from a subject. This subject sees the stimuli in the following order:
CSplusUS, CSplusUS, CSminus, CSplus, CSplusUS, CSplus, CSminus, CSminus, CSplus, CSplusUS, CSminus, CSplus, CSminus, CSminus, CSplusUS, CSplus, CSplus, CSminus, CSplusUS, CSplus, CSplus, CSminus.
The trial in question being #22, which is a CS- for this subject.
I appreciate the help!
Eric
Hi Eric
my first guess is that this happens because there is not much data left after the last event (3 s). Specifying a large response amplitude has comparatively little impact on the model fit.
However, this behaviour should still not happen, and I'll investigate whether it can be avoided. As a quick fix, if you happen to have more data at the end, then don't cut it away.
Best
Dominik
Hi Eric
I found the solution to the problem. There were some conditions under which the priors & starting values for the estimation of the amplitude could be inappropriately large. In the absence of meaningful data (as in your last trial), the estimation will fall back to this prior.
I have fixed this bug, and the amplitude estimates for the last trial will be in the same range as the others. Note however that they will still not be interpretable and just reflect the amplitude from the response averaged across all trials, because there is not enough data to estimate the amplitude properly.
To get access to the bugfix, update from the code repository.
Hope this helps
Dominik