Hi AFNI experts,
My experiment is a slow event related design, the experiment starts with a second of sound, then the next three seconds are the task time of event, and finally there is a time interval of 10 seconds between events, all the events are like this(1+3+10=14s). So what is the best regress basis of slow_event related design and how should I set up my regress basis, block or gam or tent?
Best regards,
QianqianWU
What is the time interval of interest for you: a second of sound, the three seconds of the task time of event, or the time interval of 10 seconds between events?
Gang Chen
Dear Gang
Thans for your relply. The time interval that I'am interested in is the three seconds of the task time of event. After reaing many case in AFNI discuss,maybe I shoud chose TENT(0,14,8) as the best regress basis, cause a week before,I used the GAM as the regress basis,that AFNI told me there's a lot of colinearity warnings.
Qianqian Wu
Using the tent basis function would definitely be more accurate in capturing the hemodynamic response profile. However, if the gamma basis function caused a collinearity problem, you might encounter a similar issue with the tent basis function. How many task conditions are in the experimental design? Is the interval between events fixed at 10 seconds? Could you share the stimulus timing information?
Gang Chen
Hi Gang
Thanks for your answer. There are eight task conditions in my experiment (ie,. OpenLeftSmall,OpenLeftLarge,OpenRightSmall,OpenRightLarge
GraspLeftSmall,GraspLeftLarge,GraspRightSmall,OpenRightLarge)，
and the interval between events fixed at 10 seconds.We took 8 runs, randomly repeated each of the two paired conditions (for example: OpenLeftSmall-OpenLeftLarge; OpenRightSmall-OpenRightLarge; GraspLeftSmall-GraspLeftLarge; GraspRightSmall-OpenRightLarge) 25 times for each of the two runs.
Each event consists 1s(cue)+3s(haptical event)+10s(the interval between events).The timing of each stimulus presentation corresponded as follows：9 23 37 51 65 79 93 107 121 135 149 163 177 191 205 219 233 247 261 275 289 303 317 331 345. 9 represents the first eight seconds of each run plus the time of stimulus presentation，subsequent stimulus times were incremented by 14 seconds one by one (i.e., the time for each event).
And finally the collinearity problem prompted by afni
Have you already acquired the data? If not, consider spacing the events apart somewhat randomly, with a mix of 14s, 18s, 22s, 26s (for example). This will help with collinearity -- which happens strongly when the hemodynamic responses from one event mix up with the following event in exactly the same time relationship in every case.
Hi~
Thanks for your suggestion! In fact, I'm only acquired the data from one subject right now and would like to do some preprocessing steps first. Your suggestion will be carefully considered, thank you very much!
Now we are using TENT(0,14,7) as regress basis and there is no error reported about collinearity problem, but it prompts as follows
Number of time points: 1432(before censor); 1418 (after)
Number of parameters: 1816 [80 baseline ; 1736 signal]
ERROR: Regression model has too many parameters for dataset length :
FATAL ERRoR: 3dDeconvolve dies: Insufficient data (1418)for estimating 1816 parameters
To avoid the collinearity problem, it is very important to vary the inter-trial interval.
Does each condition have 25 trials? Are you analyze each trial separately? If your TR is 2 seconds, I suggest using TENTzero(0,14,8).
Gang Chen
Hi Gang~
Thanks for your answer!
Each condition have 25 trials and I will analyze each trial separately.
I will try the new regress basis that you suggested, btw, I have a small question. For TENT(b,c,n)，does total time(14s) divided by TR(2s) equals the third parameter minus 1(n-1). I used TENT(0,14,7) as regress basis yesterday, and as I reply above, may be wrong.
And finally, I would like to ask if it is correct for me to add ''_IM'' after "stim_times" ? Or is it not necessary?
Qianqian Wu
What is your analytical goal: estimating the hemodynamic response at the condition level or the trial level? Additionally, could you share the 3dDeconvolve script that led to the collinearity issue?
Gang Chen
Hi Gang
Our goal was to average the responses for the condition level and finally to compare between conditions.
As you suggested, there were no further collinearity problems after using TENTzero(0,14,8). But 3dDeconvolve.err reported an error as follows：
*+ WARNING: Smallest FDR q [4 OpenLeftSmall#1_Tstat] = 0.424979 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [8 OpenLeftSmall#3_Tstat] = 0.103205 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [17 OpenRightSmall#1_Tstat] = 0.999899 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [19 OpenRightSmall#2_Tstat] = 0.9999 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [21 OpenRightSmall#3_Tstat] = 0.106166 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [23 OpenRightSmall#4_Tstat] = 0.110849 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [30 OpenLeftLarge#1_Tstat] = 0.263192 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [36 OpenLeftLarge#4_Tstat] = 0.158431 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [43 OpenRightLarge#1_Tstat] = 0.349963 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [45 OpenRightLarge#2_Tstat] = 0.124988 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [47 OpenRightLarge#3_Tstat] = 0.12058 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [58 GraspLeftSmall#2_Tstat] = 0.158148 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [69 GraspRightSmall#1_Tstat] = 0.999886 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [73 GraspRightSmall#3_Tstat] = 0.999898 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [75 GraspRightSmall#4_Tstat] = 0.999895 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [77 GraspRightSmall#5_Tstat] = 0.711281 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [86 GraspLeftLarge#3_Tstat] = 0.99989 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [90 GraspLeftLarge#5_Tstat] = 0.737431 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [95 GraspRightLarge#1_Tstat] = 0.119459 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [97 GraspRightLarge#2_Tstat] = 0.999898 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [99 GraspRightLarge#3_Tstat] = 0.999899 ==> few true single voxel detections
*+ WARNING: Smallest FDR q [103 GraspRightLarge#5_Tstat] = 0.133745 ==> few true single voxel detections
Qianqian Wu
You can disregard those warnings. Instead, focus on verifying whether the estimated hemodynamic responses for those conditions appear reasonable and check for strong statistical evidence in the expected brain regions.
Gang Chen
Maybe that are warings and not errors, I will refer to some of the questions and answers from others
OK, thanks for your reply~ 
Hi Gang~
How are you? I used TENTzero(0,14,8) as my regress basis and I expected it to be fine, but now I have the following error. How do I solve a problem like this？Also, this is my 3dDeconvolve script. Maybe I should not have used regress_stims_type_IM, but some other like AM1? AM2? or default times？
errors:
++ Number of time points: 1432 (before censor);1418 (after)+ Number of parameters: 1568 [80 baseline ; 1488 signall** ERROR: Regression model has too many parameters for dataset length :(** FATAL ERROR: 3dDeconvolve dies: Insufficient data (1418) for estimating 1568 parameters** Program compile date = Aug 22 2023
3dDeconvolve script:
Qianqian Wu
Qianqian Wu,
Could you provide some background information about your experimental design? What is the analytical goal? Why are you estimating trial-level BOLD responses? Additionally, what trial-level modulation variable are you considering?
Gang Chen
Hi Gang Chen~
[Could you provide some background information about your experimental design? ]
First, the purpose of our experiment was to investigate how different object positions and object sizes were represented in the haptic condition, so there were two factors in the experiment: object location(left, right), and object size (large, small), and it is worth noting that the subjects grasped only with their right hand during the experiment.
Second, in order to make the grasping process cleaner, we set up another control condition, i.e., the hand was open, which did not present the real object, but only mimicked the grasping action, so this was the third experimental condition (type of grasping: gras, open).
[What is the analytical goal?]
The purpose of our study was to see if there were differences in activation in subjects under different experimental conditions. Generally, since each experimental condition was repeated 25 times, we would like to know if it is feasible to compare subjects' results between conditions after averaging these 25 repetitions of trials? Of course, in addition to this, we would like to use the MVPA approach to see if we can decode areas in the cerebral cortex regarding tactile location and size.
[Why are you estimating trial-level BOLD responses?]
In fact, we considered the question, do we need to look at the results at trial-level? Maybe this is needed for categorization, but looking at the activation of the BOLD signal can we not ignore it?
[Additionally, what trial-level modulation variable are you considering?]
The trial-level modulation variable were object location,object size and type of grasping.
As a beginner I have so much trouble with this so thank you so much for your help!
Qianqian Wu
we would like to know if it is feasible to compare subjects' results between conditions after averaging these 25 repetitions of trials?
To address this question, estimating trial-level BOLD responses is unnecessary. Instead, you can focus on condition-level effects.
The trial-level modulation variable were object location,object size and type of grasping.
I assume that you already code each of the eight combinations separately. If that's the case, I don't see the need for adding another layer of modulation. Let me know if I'm missing something.
we would like to use the MVPA approach to see if we can decode areas in the cerebral cortex regarding tactile location and size.
Is this why you're estimating trial-level BOLD responses? If so, you might need to use a fixed-shape HRF rather than estimating the hemodynamic response shape with the tent basis function.
Gang Chen
Hi Gang Chen~
Thanks for your reply.
I think with your guidance I understand what analysis I should need to do.
First of all, we are mainly interested in understanding the activation of each of the 8 conditions and differences between conditions, so here we need to analy condition-level BOLD responses.
Then for decoding goal, we should estimate trial-level BOLD responses, so in this step of the analysis, we should use another regression function like BLOCK or GAM basis function.(?)
I'm not sure if the above are the right way to think about it, and I'd appreciate your corrections.
I'm not sure if the above are the right way to think about it, and I'd appreciate your corrections.
Yes, that sounds correct. For trial-level estimation, an alternative approach is to use GLMsingle.
Gang Chen