Hi Guruji,

We're playing with some options for noise correction of resting fMRI data using afni_proc.

The problem goes something like this:

Number of TRs: 145
TR = 2 seconds

We want to eliminate contributions to the signal with frequency greater than 1/6 = 0.167. Ultimately, we'll also probably want to get rid of frequencies lower than 1/100 but we'd like to do that with -polort when we run our seed time-series correlation analysis using 3dDeconvolve; our reason for doing this is that we occasionally see meaningful correlations among select nodes of different intrinsic networks at very low frequencies.

To accomplish this, we've included -regress_bandpass 0.001 0.167 in our afni_proc script--where the lowpass is right where we want it and the highpass is less than the lowest possible frequency in our data. Two questions emerge:

1. Is this okay?
2. The resulting (1 - ((lowpass - highpass) / nyquist)) X TRs = 46 bandpass regressors seem like a lot of degrees of freedom to lose. I *think* I get your rationale for this, which is that we're throwing away the df by considering only a certain range of the data so why not estimate the unwanted signal as well as possible. From another perspective, though, I wonder if we can keep some of those dfs to be used for some of our high-motion subjects who have a lot of TRs censored. Getting outside of my comfort zone on this, though, so just tell me if what I propose is irrational...

Thanks!

Paul