Hi Andrew,

Interpolating data to fill in censored TRs does
not seem like a good approach to me, especially
since we have known how to censor for a long time.
I guess it is to allow use of the FFT, but it also
means having to keep track of those many lost
degrees of freedom. Hopefully that is done. Plus,
one then has to make sure and apply band passing to
the other regressors.

Doing it all in one model seems much cleaner.

Regarding point 3, why would you not do all of the
pre-processing in AFNI? That is to ask, what would
you want to do that is or does not seem possible
using afni_proc.py? Or perhaps you just want to do
it all yourself, a noble but time consuming task.

If you just want to work through the steps, doing
it in parallel with an existing proc script (and
results for comparison) is a nice way to go.

One thing to note about the AFNI processing is that
the final residual time series is the same length
as the input, and therefore censored TRs are zero
volumes. That does not affect a correlation
analysis, but it might affect other types. See
the proc script for how to easily extract only the
non-censored TRs, if you want to.

Regarding the P.S., yes. There should not be much
difference in speed between the censoring approaches,
though deleting TRs should be computationally faster
than adding regressors. But it might also be a bit
more difficult to program. Maybe. I would not
ponder this much. Call them equivalent.

- rick