GLMsingle output and design matrix

[danieljanko98]

Hello all,
I am trying to run GLMsingle on data from my experiment and encountered a baffling outcome. First, my experiment is an associative learning task and consists of 3 runs. in the 1st run, the subjects learn the associations between images and their locations on a playing board (AB, CD etc.). In the second and third run, they are asked to recall the associations. First, a cue image is presented and after they've selected the location on the board, where they think the cue image is located, the card is flipped and they receive a feedback. Feedback can be either confirmatory (they found the card) or update feedback (two cards swapped their positions and they have to update their knowledge about the board).

In the first instance of my model, I was primarily interested in the following four conditions - learning, recall, confirmatory feedback, and update feedback. I used the onsets of these events in my design and transformed then into TRs and encoded to the design matrices. I got the following output

Based on the first plot, the peak is at 30 seconds after onset, is my interpretation correct? This confuses me because the 30 second HRF does not seem to have been tested (based on the plot below it). When I plot the betas and other metrics of the model, it also does no look right (especially FRAC values) -

Any ideas what could have gone wrong here? I guess I can improve it by passing a mask of the whole brain to GLMsingle so that the outside of the brain does not get processed at all?

My second question is how to set up the design when I want to test the following question. I am interested in looking at how the multivoxel patterns, associated with each item-location pair (AB), change when this pair undergoes and update (AB becomes AD). Should I treat AB pair as condition 1 and AD pair as condition 2? Meaning that when the AB is valid, the recall of AB would be condition 1A and confirmatory feedback of AB would be condition 1B. Once the swap happens, the new AD becomes condition 2 and cueing of the pair is condition 2A and confirmatory feedback is condition 2B. Would it make sense to code it like this?

Thank you in advance for your help!

[kendrickkay]

Some thoughts: - In the runwiseFIR.png, I agree that it looks weird/bad/puzzling. Typically, one should check and doublecheck that the experiment presentation and timing is exactly what you intend, and to check that your preparation of the design matrix is exact and accurate. Another set of issues to consider is whether your particular coding of the design matrix is an appropriate one for the types of brain responses you think you are eliciting. Another thing that might be useful as a diagnostic is to deliberately mess up the design matrix and then see how the outcome measures look. If the current results look indistiguishable from a deliberately messed up case (e.g. also, you could scramble data / use resting data / etc.), then that narrows the possible problems. Another, more drastic route would be to start with a simpler paradigm and make sure that the whole pipeline from raw data to GLMsingle outcomes makes sense, before moving on to more complicated cases. - Masking the raw data won't really fundamentally change things, so I wouldn't worry about that. - Regarding the second question, practically, you should figure things out satisfactorily for the basic / easy case before worrying about these additional issues. But in general, the decision of how to code your design matrix is one that doesn't really interact/impinge on how GLMsingle works (in the sense that whether a design matrix is coded correctly/appropriately is a question that is to be resolved by the user, prior to worrying about how to estimate the associated model (which is what GLMsingle does)). Hope this helps.

…

On Dec 4, 2025, at 2:38 AM, Daniel Janko ***@***.***> wrote: Hello all, I am trying to run GLMsingle on data from my experiment and encountered a baffling outcome. First, my experiment is an associative learning task and consists of 3 runs. in the 1st run, the subjects learn the associations between images and their locations on a playing board (AB, CD etc.). In the second and third run, they are asked to recall the associations. First, a cue image is presented and after they've selected the location on the board, where they think the cue image is located, the card is flipped and they receive a feedback. Feedback can be either confirmatory (they found the card) or update feedback (two cards swapped their positions and they have to update their knowledge about the board). In the first instance of my model, I was primarily interested in the following four conditions - learning, recall, confirmatory feedback, and update feedback. I used the onsets of these events in my design and transformed then into TRs and encoded to the design matrices. I got the following output runwiseFIR.png (view on web) <https://github.com/user-attachments/assets/9cf79af4-e30f-407f-a9f9-e77e7c5f5873> Based on the first plot, the peak is at 30 seconds after onset, is my interpretation correct? This confuses me because the 30 second HRF does not seem to have been tested (based on the plot below it). When I plot the betas and other metrics of the model, it also does no look right (especially FRAC values) - Screenshot.2025-12-04.at.09.34.15.png (view on web) <https://github.com/user-attachments/assets/736f212f-0141-4452-9387-31bc19564fe4> Any ideas what could have gone wrong here? I guess I can improve it by passing a mask of the whole brain to GLMsingle so that the outside of the brain does not get processed at all? My second question is how to set up the design when I want to test the following question. I am interested in looking at how the multivoxel patterns, associated with each item-location pair (AB), change when this pair undergoes and update (AB becomes AD). Should I treat AB pair as condition 1 and AD pair as condition 2? Meaning that when the AB is valid, the recall of AB would be condition 1A and confirmatory feedback of AB would be condition 1B. Once the swap happens, the new AD becomes condition 2 and cueing of the pair is condition 2A and confirmatory feedback is condition 2B. Would it make sense to code it like this? Thank you in advance for your help! — Reply to this email directly, view it on GitHub <#197>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAU2DEWISGVX2SZBVVOIEJ3377XHPAVCNFSM6AAAAACOAJFWHOVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZZGIYDOMBQGU>. You are receiving this because you are subscribed to this thread.