Hi Janina,
It is great that you are reviewing those threads!
Censoring is done because motion causes large spikes
in the time series that will not be well modeled by
typical regressors, and large spikes are problematic
when finding a least squares solution.
If bandpassing is performed in the presence of big
spikes that have not been censored out, they can
end up ringing throughout the time series, leaving
smaller echoes of the spikes in the result. So any
subsequent censoring can no longer fully remove the
spike from the time series.
Your concern also applies, and is a different problem.
Yes, the regressors would also have to be bandpassed,
or else the result would end up putting those unwanted
frequencies back in. But that problem is solvable,
as long as people are aware of it. The previous one
is probably not (well the solution is to do it via
linear regression, not an FFT).
Yes, bandpassing is exactly projecting out those
unwanted frequencies. It can be done via an FFT or
via regression, as afni_proc.py does it. And every
signal that is projected out costs one degree of
freedom, which are probably lost due to the separate
steps.
I do not know how people outside the AFNI community
account for this. I suspect that for the most part,
they do not. As a consequence, it is likely that
many papers get published with subject results that
are akin to having negative degrees of freedom, i.e.
they are based on garbage.
Even aside from negative (or just accounting for the)
degrees of freedom, the problem of bandpassing as a
separate step is likely not well handled.
- rick