Hi Rick,
Quote
1. Yes, those bandpassing parameters should be okay. You
are probably using a quadratic or possibly even cubic polort.
Note that using 0.01 instead would probably cost another 4 or
8 regressors.
Okay. Thanks for the affirmation.
Quote
2. I agree that it is a lot of regressors to lose, and you
are giving up fewer than most because 0.167 is a bigger
cutoff than most use. This is why I believe bandpassing
is too high a penalty to pay in general. There are more
efficient ways to remove likely noise.
Are you thinking about procedures like ricor? We often do apply this but there's no guarantee that it factors out all the high-frequency noise. I've liked the idea of just proceeding with the assumption that anything in the BOLD signal that completes a cycle in 6 seconds is likely non-neural in nature and should just be cut out. We've been doing this with a regression-based approach because we also want to censor bad TRs. What I might try in order to preserve some sensitivity is to run 3dBandpass on the data before even introducing them in the afni_proc context.
Quote
Essentially, bandpassing projects the data into a 99
(= 145-46) dimensional sub-space. It is like having 99
TRs left to model with. Modeling out another 12 for
motion, 3 for the polort and 50 for censoring would drop
it down to 34 remaining df to correlate with.
Yup. Then imagine adding ricor and rvhrcor regressors to the mix. It's a tough balancing act.
Quote
What is your motion censoring threshold?
It's 0.2 mm and we disqualify subjects with more than 20% of TRs lost to motion censoring. Do you know of any theoretical or empirical means for making the decision to cut off at a certain motion threshold, like 0.2 mm? Like something that shows loss of useful signal as a function of per-TR motion?
Big thanks!
Paul