Using GLMsingle in a Stop-Signal Task: Trial Variability vs. Repetition Structure

[gaozhy]

Hello,

I am planning to apply GLMsingle to a classical stop-signal task (see figure):

Go trials: fixation (0.5 s), then a left/right green arrow requiring a speeded response.

Stop trials: fixation (0.5 s), then a green arrow, followed after a variable SSD by a red arrow instructing participants to withhold their response.

The response window is 1.5 s, and ITI is ~1.5 s. Each participant completed two runs in the scanner.

The main goal of my project is to examine trial-by-trial variability in cognitive control.

My concern is that if I assign all trials of the same type (e.g., all go trials, or all go trials with the same arrow direction) to one condition, GLMsingle will treat them as repetitions. While this is correct perceptually, I worry it could remove meaningful trial-to-trial variability, which is central to my study. The eLife paper suggests this may not be an issue, but I’d like to confirm whether that still holds when the goal is to analyze trial-by-trial variability in cognitive control.

One idea is a hybrid approach: use only a small subset of go trials from each run as repetitions (to enable denoising/ridge regression), while modeling all other go, successful stop, and failed stop trials as single-trial regressors/conditions. The repeated subset would be excluded from downstream analyses.

Do you think this setup is reasonable, or is my concern about repetitions reducing trial-level variability unfounded?

Thank you very much for your guidance!

[kendrickkay]

"While this is correct perceptually, I worry it could remove meaningful trial-to-trial variability, which is central to my study. The eLife paper suggests this may not be an issue, but I’d like to confirm whether that still holds when the goal is to analyze trial-by-trial variability in cognitive control."

Having read your description, I would say that it should not be a problem. As you mention, the use of the repetitions is to set the hyperparameters for glmdenoise and RR. Ultimately, you receive single trial estimates, and these estimates should still have any real/neural trial to trial variability in the data. For more information, see https://glmsingle.readthedocs.io/en/latest/wiki.html#in-glmsingle-the-glmdenoise-and-ridge-regression-rr-components-of-the-method-require-experimental-conditions-to-repeat-across-runs-how-should-i-think-about-whether-this-is-appropriate-for-my-experiment

The hybrid approach that you mention is sensible, but I don't think necessary to pursue

[gaozhy]

Thanks very much for your quick response and your confirmation, that's very helpful! I'll proceed with the method and skip the hybrid approach.