Questions about setting up a design matrix and stimuli repetition

[AndyJHKim]

Hi!

I’ve really enjoyed reading GLMsingle paper, and I’m now using GLMsingle for my own project. However, while working with it, I’ve run into a few issues and wanted to check whether I’m setting up the design matrix correctly.

About my experiment:
I’m running a lexical decision task with three conditions:
Words (e.g., apple, bang) - press left button
Pseudowords (e.g., appli, beng) - press right button
Nogo (e.g., xxxx) - no response

There are 6 runs.
Each word and pseudoword appears exactly once in the entire experiment, whereas nogo stimuli appear multiple times.

1. Given this structure, I’m unsure what the correct form of the design matrix should be for GLMsingle. For example, suppose we have two conditions (A, B) and three stimuli (A-1, A-2, B-1), where B-1 appears twice, with onsets at TRs 0, 2, 1, and 3. Should the design matrix be:
   Condition-level columns (A vs B)
   [1, 0]
   [0, 1]
   [1, 0]
   [0, 1]
   Or
   Trial-level columns (A-1, A-2, B-1)
   [1, 0, 0]
   [0, 0, 1]
   [0, 1, 0]
   [0, 0, 1]
   Or since GLMsingle internally flatten everything to single-trial (I mean, single-occurrence?) regressors, it doesn’t matter?

2. I want to confirm whether having only one occurrence (no repetition) of most stimuli is sufficient for cross-validation. In my experiment, most stimuli appear only once throughout the whole experiment. Is this acceptable for running the analysis? I believe it should be fine, because when I examined example datasets and code, I also noticed that some stimuli appeared only once, but I want to double-check to be sure. Do you require all stimuli to be repeated in different runs?

Thanks in advance!

[kendrickkay]

A few thoughts:

• Are you coding Nogo as a condition? If you are (and if that's your intention), this will help give GLMsingle trials to cross-validate on.
• Regarding your (1), either way could be fine. For what you call "condition level" you are telling GLMsingle that you generally expect responses to be similar (cross-validate) at the condition level. For what you call "trial level" you are telling GLMsingle something more conservative, namely that you expect responses to be similar only for the exact same stimulus. The downside of this approach is that you may not have enough trial repetitions for GLMsingle to use.
• It may very well be that regardless of the approach, the single-trial estimates you get from GLMsingle may be quite similar. See this issue here: 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
• Regarding your (2), if you code conservatively and most of your conditions are specified as happening only once, then the downside is that you need at least some repetitions for GLMsingle to use (to tune the GLMdenoise and ridge regression components). Hence, that's why I asked about the Nogo condition and whether you are thinking of that as a condition that repeats.

[AndyJHKim]

Thanks so much for your quick and detailed reply!

Yes, I am coding Nogo as a condition. My reasoning was that Nogo trials are still stimuli, so it made sense to include them as regressors. In my case, even the Nogo category consists of 5 different stimuli (the number of “x” characters varies from 4 to 8). I was unsure whether I should model them as 5 separate regressors or collapse them into a single Nogo regressor.

Based on your explanation, it seems that the results might not differ dramatically either way, which is reassuring. Ideally, I would prefer not to run GLMsingle multiple times for different hypotheses, so it’s helpful to know that the choice of grouping might not drastically change the outcome.

Given this, I think I will start by modeling each stimulus separately (i.e., one regressor per stimulus), since I want to test several trial-level hypotheses beyond the broad categories of word vs. pseudoword vs. Nogo. If time and computational resources permit, I may later compare this with a more compact model using only the three category-level regressors.

Thanks again for your helpful guidance!

[kendrickkay]

Yes, I am coding Nogo as a condition. My reasoning was that Nogo trials are still stimuli, so it made sense to include them as regressors. In my case, even the Nogo category consists of 5 different stimuli (the number of “x” characters varies from 4 to 8). I was unsure whether I should model them as 5 separate regressors or collapse them into a single Nogo regressor.

The considerations are similar to the prior issue that I commented on. It likely won't make a big difference.

Given this, I think I will start by modeling each stimulus separately (i.e., one regressor per stimulus), since I want to test several trial-level hypotheses beyond the broad categories of word vs. pseudoword vs. Nogo. If time and computational resources permit, I may later compare this with a more compact model using only the three category-level regressors.

I understand the logic you are stating here. But note that GLMsingle ultimately gives you a separate estimate for every single trial in your experiment and does so in a way that shouldn't greatly depend on the design matrix specification you provide it (including the choice of what to group together that you have been deliberating about). So, you could try even the three category-level regressor approach

[AndyJHKim]

I see. Thanks!!