Looking for advice on a GLM

AFNI version info (afni -ver): Precompiled binary linux_openmp_64: May 1 2024 (Version AFNI_24.1.06 'Publius Septimius Geta')

I have an encoding task and a retrieval task. Encoding is designed to be ran during a scan (i.e. fixed stimulus onset and duration), while retrieval is designed to be ran outside of a scan (i.e. task is self-paced).

Due to some unfortunate miscommunication, a few subjects were scanned while they did the retrieval task. But! Looking at the timing for one subject, it looks like all stimuli are at least 1 TR apart. Also, retrieval task was multiple-choice, i.e. each stimuli is either "correctly" or "incorrectly" guessed. It looks like the scan was ran continuously, so there is considerable dwell-time at the end. With full acknowledgement that this probably won't produce anything too spectacular, how should I model the retrieval task? Would it even be meaningful to do so?

Main way I was thinking was modeling "incorrect" and "correct" as two separate -stim_times_AM2 and using dmUBLOCK, but I'm open to suggestions!

(full disclosure: this is more for my own curiosity / enrichment than anything else ^_^)

When you refer to "event," are you indicating a combination of encoding followed by retrieval? How much variability is there in the interval between consecutive events? Similarly, how much interval variability exists between an encoding and a retrieval? Given that each retrieval is classified as either "correct" or "incorrect," what are your primary effects of interest: the two retrieval types or the events themselves? In other words, are you looking to distinguish the BOLD response between encoding and retrieval phases?

Gang Chen

Relevant info: TR is 1.5 seconds.

Quick clarification: "Encoding" and "Retrieval" are two separate tasks. "Encoding" was designed to be ran during a scan, while "Retrieval" was designed to be ran outside of a scan (i.e. task was self-paced). Due to miscommunication, Retrieval was mistakenly ran in-scanner for a few subjects.
By "event", I meant "stimulus presentation". Edited post to reflect this.

Encoding stimuli had fixed onset and duration (5s), with variable jitter (1-4s).
Retrieval task was self-paced, i.e. stimuli were presented and remained on screen until a subject provided a response. The next stimulus was immediately presented after a subject response (no delay time, I believe).

So, Retrieval stimuli duration (in this case, difference between onset times of n+1 and n stimuli) is highly variable. E.g. for a few subjects:

mean sd min max
7.4 4.2 1.9 18.8
5.6 4.9 1.5 25.0
11.3 10.6 3.2 60.8

I'd naively say my effect of interest is in retrieval type, i.e. between "correct" and "incorrect" retrievals.

If I understand correctly, each retrieval occurs consecutively without a gap, even though a resting period is included at the end. If that's the case, the conventional modeling approach is likely not going to work well, including the -stim_times_AM2 option with dmUBLOCK.

However, it might be worth exploring a data-driven approach. For instance, if the entire retrieval period spans 150 seconds, you could use the first onset time in conjunction with something like TENT(0,165,111). Then, examine the estimated BOLD response over the full period and annotate it with the retrieval onset times. This might help you identify any patterns differentiating "correct" and "incorrect" retrievals.

Gang Chen