Dataset for nilearn tutorial

[htwangtw]

Currently we are directly converting the nilearn examples directly using nilearn.datasets
It would be good to expand it to a full analysis.
We will need to find a preprocessed dataset on task-fMRI for this.

[htwangtw]

Let's use ds000001 @yibeichan
It's only 16 subjects if we want to run the full process, and we can try to run a smaller subset (i.e. 5 subjects).

[yibeichan]

sounds good! I can experiment with 5 subs first

[yibeichan]

hello @htwangtw @effigies , this dataset has 3 runs for each subject, since our goal is a 2-level glm. We have two choices:

1. concatenate three runs into one, run the first-level on this concatenated run, then do the second level
2. do the first level for each run, average copes/varcopes across three runs, then do the second level.

I prefer the first solution, because the second one sounds like a three-level glm.

[htwangtw]

2 is the correct way and how the data was analysed in the original paper! I don't think it's a bad thing to make it three-level. I would suggest to have a look at the original paper.

[yibeichan]

Yes, the original paper used FSL, so they did three-level GLM.
I understand that 2 is correct. But why 1 is not?
GLM is averaging across trials. So as long as we concatenate events.tsv, preproc_bold.nii.gz and confounds.tsv in the same way, then GLM on the concatenated run should be the same as GLM on each run and average?
Or because of the conjunction of each two runs will cause some differences (e.g., in terms of hrf)?
Sorry, I only did GLM 2-3 times; I am probably wrong

[htwangtw]

It's okay to concatenate runs but you will have to make sure the signal are normalised, so they are comparing with the same baseline. The safest way is to do each run separately, and then combine the statistical maps across run.

[yibeichan]

Ah I see, I'll go for option 2 then, much safer. Thank you!