@stim_analyze and make_random_timing.py

AFNI Message Board
1

AFNI version info (afni -ver):
Precompiled binary linux_openmp_64: Feb 25 2025 (Version AFNI_25.0.07 'Severus Alexander')
I'm using @stim_analyze with make_random_timing.py to evaluate stimulus presentation order efficiency. I have a few questions:

  1. What is the difference the search patterns LC and norm. std.? Are there situations where one is preferred over the other?
  2. The event durations will vary based on subject response times. Currently, I'm setting event durations to the maximum possible. Will it be a problem when the durations are actually slightly different from what I'm using to evaluate design efficiency?
  3. Is there a maximum acceptable LC/norm std value? For example, if my most efficient design (out of 10,000 iterations) has n. std. = 0.410400 and there are no matrix warnings and all matrix conditions are labeled VERY GOOD, is that still too large a normalized standard deviation?

Thanks!

Phil

2

Hi Phil,

  1. The 'norm. std.' output lines refer to the individual regressors, while the LC output lines refer to contrasts. So to evaluate the sum of contrasts, use the LC search pattern.
  2. This cannot be done perfectly, so just do it reasonably. I might be inclined to use the expected average response time (so events/time averages out), and maybe even let the duration vary (probably uniformly, as there is no option for a normal distribution, for example). In this case, I guess you are really just getting a randomized list of events, rather than applying events and onset times, is that right?
  3. Those values will be based on the design, so we do not really evaluate them directly. Maybe @Gang has newer thoughts on this.

-rick

3

Hi Phil,

Below are some considerations that emerged from a recent reflection on fMRI experimental design. I’d be very interested in hearing any feedback on how these ideas align with (or diverge from) real-world practice:

  • Stimulus timing need not be synchronized to acquisition grids and, when allowed to vary, can improve estimability.
  • Varying stimulus sequences across participants is generally preferable.
  • Trial-level sampling can rival participant sample size in its importance for inference.
  • Interaction effects typically require disproportionately greater sampling.
  • High cross-condition correlation may, somewhat counterintuitively, benefit group-level inference.
  • More flexible HRF modeling can improve both efficiency and robustness.
  • Stimulus attributes (e.g., trial duration) can improve estimation precision, whereas behavioral covariates (e.g., reaction time) may introduce bias.
  • Unequal group sizes may be statistically optimal under heterogeneous variance.

One implication that seems directly relevant to your case is that packing in as many trials as feasible may be advantageous. Concerns about high cross-condition correlation may be overstated in common practice.

Gang Chen

4

Thanks for the clarification about LC vs. norm. .std. I was using norm. std., but it sounds like LC would be better.

What motivated this is concern that the behavioral effects observed outside the scanner will weaken or disappear if we deviate too much from that paradigm's fixed ITI of 500 ms. I'm currently running these simulations with an ITI that varies between 0 and 5,000 ms. The real question is whether 5 seconds is too short for a maximum ITI? Could I go even shorter than that? I was hoping these 3dDeconvolve simulations with no data would give a clear answer to that question.

5

Thanks for the suggestions, Gang! I will take take as many of these as is feasible into consideration--scan time is an issue.

This is how I normally approach this and I would typically just randomize the ITI for each subject, but I was concerned that with a shorter maximum ITI than I would normally use (5 seconds vs. around 12 seconds), I might be better off evaluating design efficiency. I guess I could use the highest ranking sequences, with each subject getting a different one, assuming some ITIs don't need to be longer than 5 seconds (or less).

Phil

6

Hi Phil,

I was concerned that with a shorter maximum ITI than I would normally use (5 seconds vs. around 12 seconds), I might be better off evaluating design efficiency. I guess I could use the highest ranking sequences, with each subject getting a different one, assuming some ITIs don't need to be longer than 5 seconds (or less).

In this context, two opposing factors jointly determine estimation precision (or efficiency): the degree of correlation between a condition regressor and other regressors, and the number of trials. For a fixed scan duration, increasing the number of trials typically raises regressor correlations (thereby reducing estimation precision) while simultaneously improving precision by providing more observations for the condition of interest. In practice, the gain from increased trial numbers generally outweighs the loss due to higher collinearity, which is why shorter inter-trial intervals are usually preferable.

Gang Chen