Event regressors

[edickie]

Event regressors: EA video osnets and durations Circles videos onsets and duration EA button presses (Dur = 0) Circles button (Dur=0) parametric modulators or EA (edited) 1:50 Regressors: mf.CSF mf.WhiteMatter fd mf.X mf.Y mf.X mf.RotX mf.RotY mf.RotZ and then diff for: mf.X mf.Y mf.X mf.RotX mf.RotY mf.RotZ

[loliver4]

The smoothed files Thomas made and the tsv files are located in /projects/colin/SPINS/EA_test_spm (note: the tsvs are also located in each participant's data/nii dir). The files *_desc-preproc_Atlas_s6.nii can be used as inputs into AFNI or SPM, as best as Colin understands.

[loliver4]

Discussed the options of including parametric modulation for button presses (which has been done in the past for EA by other groups, Joe, and myself), or modeling button presses as events. For test purposes at least, we are modeling the presses as events with a duration of 0, though we may want to look at both options for the actual pipeline.

We are also opting to model both button presses during EA and the Circles Task, especially for those who want to look at residual activation (background connectivity).

[loliver4]

Also discussed selecting spm for hemodynamic model choice in nilearn, for test purposes. Need to check how this aligns with AFNI.

[loliver4]

Will also want to further discuss the inclusion of a polynomial (degree?) for baseline detrending - used to account for drift (voxel-wise correlated drifting), or alternative options.

[loliver4]

Re modeling of button presses: Colin has noted a lot of very close button presses, especially in circles where the color often changes a few shades at once. He was thinking of collapsing presses that were one second apart, so we aren't modeling a series of consecutive events, which results in a very large modeled hemodynamic response.

Lindsay and Erin noted that this may be why the original paper chose to use a parametric modulator.

Colin thinks the event-related modeling is still better for the button presses, as "if they press buttons, lets say, 4 times in the middle of the block, then we get discrete and specific increases in BOLD during those specific segments of the block in regions that are involved with the motor response.

It only effects the total BOLD response across the block in a very specific region/time. Fitting the parametric modulator doesn't effect the beta for the actual EA video blocks, and is a bad way to model out motor responses, as it will try and fit a general variation on the total block amplitude. When we take our residuals, if there are a few clusters of motor responses, we have failed to remove the specific motor time series, while having modified the residual amplitude of the areas where the video blocks occurred in a sort of weird and non specific way."

Currently moving forward with exporting all button presses as events, rather than collapsing across presses due to realization that it could get complicated with press durations at that point etc.

Maybe more justification for comparing the amplitude modulated vs event-related modeling of button presses for the actual pipeline.

[loliver4]

Zaki et al. justify the parametric modulation of EA button presses as follows: "We also included a regressor of no interest corresponding to the amount of affect ratings perceivers had made per minute during each video; this allowed us to control for the possibility that an increased number of ratings, and not accuracy per se, would be driving brain activity during accurate blocks."

From Harvey et al.: "To ensure that the neural correlates of EA were not confounded by the sheer number of button presses, the analysis included a parametric regressor of no interest representing the number of button presses per video."

[loliver4]

The data we should be using for testing is in /scratch/jjeyachandra/bidsify/SPINS, and both the original confound files and those with fixed.tsv suffix.

[loliver4]

AFNI doesn't let you model an event with duration of 0, so I used 1 for the button presses atm. My individual outputs are in /projects/loliver/SPINS_GLM_Test/sub-*

[loliver4]

Even after averaging EA + confounds across runs, pmods in nistats still look variably different than those from AFNI and SPM. After all this testing, it feels like we’re having to hack and trying to get nistats to do what AFNI is doing by default, so we’ve decided to move fwd with AFNI for EA only, due to the parametric modulator.

Plan to use 36p regression (6 motion, plus mean WM, CSF, global signal, and squares, derivatives, squares of derivatives for each). Will also run without GSR (32p). Will include polynomial regressors using option A in AFNI, which means the program will automatically choose a value based on the time duration D of the longest run.