Dear Rick,
Sorry for my persistant lack of understanding, but I have still difficulties with this topic. I read all I could find about it on the message board but there is still some confusion:

From: [afni.nimh.nih.gov]
You wrote:
“Censoring before bandpassing breaks the time axis in a
model that is all about signal frequencies. Censoring
after bandpassing can send echoes of would-be-censored
spikes ringing through a time series.”

--> I don’t understand the second statement completely. Is this only the case if the other regressors are not filtered the same way, thereby re-introducing noise back into the signal? Or would this concern still hold if each regressor was filtered with the same bp (eg .01-.1) before the regression?

From: [afni.nimh.nih.gov]
You wrote:
“But note that losing that many degrees of freedom is not
just a rationale, that is a real cost. You are projecting
out those high frequencies by modeling 46 regressors of no
interest. Running something like 3dBandpass would simply
hide such a fact, which is one of many reasons why we do
bandpassing via the linear regression, instead.”

--> Does that mean that bandpassing removes always the same amount of degrees of freedom, no matter whether it is done with 3dBandpass before regression or within the regression, with the only difference that in the regression it is obvious?

From: [afni.nimh.nih.gov]
You wrote:
“A discrete Fourier transformation can be greatly sped up
by using an FFT (Fast Fourier Transform) algorithm, which
takes advantage of the fact the regressors are sines and
cosines. But it is still doing the same thing: finding
amplitudes/beta weights for each component/regressor.
While an FFT implementation should be much faster than a
regression implementation, regression allows for proper
handing of censoring.”

--> So is the bandpassing step actually the same, no matter whether done separately before or within the regression, but if doing it separately it is problematic to combine it with censoring and one might be unaware about how many degrees of freedom are actually lost due to the procedure?


It seems to me that outside the afni community, bandpassing separately is still quite common. How would people account for the lost degrees of freedom in that case?
Can you recommend any articles on this topic?

Thank you very much!
Janina