Questions about event onsets not aligned to TR & design matrix construction

[LouiseRyan95]

Dear Kendrick,
I am a first-year graduate student and very new to this field, so I apologize in advance if these questions are basic. I recently started working with fMRI analysis and have encountered some issues that I would really appreciate some guidance on.

Experimental design:

• Each image is presented for 1 second
• Followed by a jittered ISI of 2–3 seconds
• TR = 1 second
• Each subject has 11 sessions
• Each session contains completely different images
• Each session is repeated three times

A practical issue I am facing is that many stimulus onsets are not aligned to the beginning of a TR, as you can see in the events.tsv file I uploaded.
test.py
sub-001_ses-102_task-s1_events.tsv

So far, I have tried two approaches:

• Resampling the TR from 1 s to 0.5 s
• Keeping the original TR (1 s) without resampling

Under these two settings, if my codes are right, the overall results are very similar. For example:

• The mean R² increases slightly from approximately 4.28% (no resampling) to 4.7% (with resampling).
• The split-half reliability is almost identical between the two approaches.

My questions:

How should stimulus onsets that are not aligned to TR boundaries be handled in this kind of design? Is temporal resampling (e.g., from 1 s to 0.5 s) a reasonable or recommended solution here?

If possible, could someone help check whether my code logic is reasonable, or point out any obvious conceptual mistakes?

I have attached:

• An example of my events.tsv
• My code

Any advice, references, or corrections would be extremely helpful. Thank you very much for your time and for maintaining this project.

Best regards,
Gengchen Ye

[kendrickkay]

"the overall results are very similar" ==> This is not surprising, and it's a good outcome.

In general, upsampling the time-series should only improve things (although the improvement may be modest). The rationale is as follows: If your specification of the design matrix is inaccurate, this is expected to lead to sub-optimal outcomes. If, by upsampling, you can specify a design matrix that is more accurate, this should lead to improvements.

"I do have a 2s TR and want to convert the onsets in seconds into the design (matlab code). Which rounding do you recommend to apply? e.g., floor(onsets/tr) +1 ? Or just round(onsets/tr)?"

Assuming that onsets are decimal times in seconds where 0 s corresponds to the first row of the design matrix (and the first fMRI volume), then round(onsets/tr)+1 is the correct thing to do. Consider a simple example. If an onset is at 1.9 s and the TR is 0.5 s, the closest volume to the onset is the fifth fMRI volume. round(1.9/.5)+1 = 5. floor(1.9/.5)+1 = 4 and that is incorrect.

"The stimulus duration is 0.5s, so should I resample the files?" ==> The issue isn't really about the duration of your stimulus, but it's more about the onsets of the events and whether they land on the onsets of the fMRI volumes.