multiple runs design,, how to account for signal decay across runs?

AFNI version info (afni -ver):

Dear, AFNI experts.

I have a 2(A1,A2)*2(B1,B2) design task which consists of 4 runs and one of my experimental variable was set to be in between runs .
For example, If I wanted to create contrast between A1>A2. I'd have to model across run1,run2>run3,run4
and the order of condition was counter balanced between subjects/

the behavioral result showed strong difference between A1, A2 (after having accounted for the order effect) however, the neural result of A1>A2 contrast is always non-significant.
I have looked over some literature stating that the design comparing over long time difference is not recommended, and I know that there must be a decay in signals between runs.
I was initially using SPM12 but have come across AFNI that it's more flexible and looked at many topics that might be related to such design.

What ways would you recommend for such design? My idea was,,

  1. is it possible to account for decaying signal effect across runs using AM regression? and what numbers am I supposed to input for each runs then?

  2. What other ways would you recommend??

Any advice would be very helpful,, thanks so much in advance

code text  # or delete if not needed

I have a 2(A1,A2)*2(B1,B2) design task which consists of 4 runs and one of my experimental variable was set to be in between runs .

To clarify, do you have four task conditions—A1B1, A1B2, A2B1, and A2B2—in your experiment? Is the order of these conditions randomized across participants, and is each condition confined to one of the four runs? How many trials are there per condition? Is the design event-related or block-based? Additionally, how are you modeling the hemodynamic responses—using a canonical HRF or through HRF estimation?

If you're considering amplitude modulation with the -stim_times_AM2 option, it could be appropriate if you assume a linear attenuation a priori. Alternatively, you might estimate trial-level responses and analyze the attenuation effect at the group level. However, if you're simply comparing A1 and A2, modeling the attenuation effect may not be essential.

Gang Chen

Hi, Gang.
First of all, thank you greatly for your reply. I wasn't expecting to get help this fast and was about to give up on learning to use AFNI ,,

and I wasn't clear on my explanation about my task design, and I apologize that made a mistake on a previous posting because what I want to see is actually contrast of A>B.

(My task is actually 2X 2 X 5 design, but I usually account for the last variable as parametric modulator or just ignore the last variable only for simplified analysis. and whether I include the last variable or do the simplified version, the result doesn't change much that contrast between A>B is always non significant....)

Anyway, to answer your question,
1) do you have four task conditions—A1B1, A1B2, A2B1, and A2B2—in your experiment?

yes, it's a four conditions task
*A1 – Disclosed, Higher Status
*A2 - Disclosed, Lower Status
*B1 - Hidden, Higher Status
*B2 - Hidden, Lower Status

2) Is the order of these conditions randomized across participants, and is each condition confined to one of the four runs? Is the design event-related or block-based?

I had 4 runs total,
in run 1,2 condition A1,A2 was randomized in event-related fashion.
in run3,4 condition B1,B2 was randomized in even-related fashion.
(my supervisor recommended a cross-over design for my experiment)

And I guess for condition A,B it was a very long block design?
I hope my diagram helps you to understand better.

Each run lasts about 8 to 10 minutes, and the response was always self-paced.
order of condition A,B was counterbalanced between subjects.

3) How many trials are there per condition?

I have 60 trials per run. (60 for each condition)

4)Additionally, how are you modeling the hemodynamic responses—using a canonical HRF or through HRF estimation?

yes, I have been using a canocinal HRF + stick function through spm.
In AFNI I used GAM function which drastically changed the results in subject level analysis. (I'm planning to ask this in another topic) My supervisor doesn't favor(?) TENT function so I can't use this in my analysis.

What is the equvalent of a canonical HRF + stick function in AFNI?

5)If you're considering amplitude modulation with the -stim_times_AM2 option, it could be appropriate if you assume a linear attenuation a priori. Alternatively, you might estimate trial-level responses and analyze the attenuation effect at the group level. However, if you're simply comparing A1 and A2, modeling the attenuation effect may not be essential.

so this means that I won't need to consider modeling attenuation across 1,2,3,4 runs instead just consider trial-wise attenuation effect.

and also If I'm assuming a linear attenuation a priori,
AFNI's defautl setting already considers this linear drift as - polort 5 setting?

Thank you greatly for your contribution and I hope you have a wonderful day,

-Saim

what I want to see is actually contrast of A>B

Does this mean you're not interested in comparing A1 and A2, or B1 and B2? If so, does that imply you don’t need to differentiate between A1 and A2 and between B1 and B2 in your model?

How were the A1/A2 and B1/B2 trials arranged? Were there jittered inter-trial intervals, or were the trials presented consecutively without gaps?

the result doesn't change much that contrast between A>B is always non significant

Were the results non-significant at both the individual and group levels? At what threshold was significance assessed?

What is the equvalent of a canonical HRF + stick function in AFNI?

Which aspects did you model with the canonical HRF, and which were modeled with stick functions in SPM?

Gang Chen

Hi, gang. thank you again for your response!

Does this mean you're not interested in comparing A1 and A2, or B1 and B2? If so, does that imply you don’t need to differentiate between A1 and A2 and between B1 and B2 in your model?

yes, I am interested but, the behavioral result is strongly different between A>B so, I wanted to see what neural mechanism were there underlying the A>B effect.
in my model, I usually model 0.5xA1+0.5xA2 - 0.5xB1-0.5xB2
for reference I post the design matrix below

How were the A1/A2 and B1/B2 trials arranged? Were there jittered inter-trial intervals, or were the trials presented consecutively without gaps?

yes they were jittered inter-trial intervals. the ITI were 1.5s ~ 3.5s and I caculated the optimal intervals from OptimizeX package. but since my task got responses from participant in self-paced time,, the interval between each trial greatly differs between subjects

Were the results non-significant at both the individual and group levels? At what threshold was significance assessed?

some individuals show great effect at subject level and others don't, which I think accounts for the non-significant group-level effect.
I suspect maybe the long term between A,B might be the problem? so, that is why I'm looking for ways to solve this problem like considering for signal decay between runs1,2,3,4..

I assessed it at p=0.001 and did a clusterwise correction

Which aspects did you model with the canonical HRF, and which were modeled with stick functions in SPM?

I'm not quite sure what this means,, but I modeled it so that it has 0 duation,, (not boxcar which considers duration of the stimuli)

Thank you so much for your comment and hope you have a wonderful day :slight_smile:

I assessed it at p=0.001 and did a clusterwise correction

The thresholding criterion you're using may be unnecessarily conservative, potentially limiting the informative value of your results. You might achieve better insights by exploring alternative visualization approaches for result reporting, as discussed here.

Additionally, the use of a canonical HRF could be reducing your detection sensitivity. Consider improving your modeling efficiency by adopting a more flexible HRF estimation approach, as outlined here.

Gang Chen

thank you very much :)