Hi, Philipp-

Like the Fourier Transform, regression modeling uses an entire time series. (In fact, the FT is a special case of linear regression modeling---just using specific regressors---which is why afni_proc.py does bandpassing as a regression, to do mathematically correct processing when wanting to both bandpass and regress a model.)

Your modified question still leads to an "it depends" answer. Baseline modeling takes place during processing, and this will generally be different between your two cases (chop then process, vs process then chop). Now, in special cases of certain time series properties, the baseline model might be essentially the same between the two cases---such as if there aren't large fluctuations of certain orders. But again, that depends on the specific time series properties and the relative size of chopping.

Re. fundamental frequencies: I think the more relevant point is the Nyquist, which depends on TR. Having more or fewer time points (for a constant Nyquist) means that your frequency spectrum is mroe or less fine, respectively, sure, but if averaging over the same band of frequencies, then the "smearing" should reduce that a bit. MRI time series are so noisy, I don't think one would want to focus in on one, specific frequency very often unless one is looking for some mechanical effect, say. Even breathing rate won't be perfectly constant, so one might estimate breathing effects over a small range of frequencies (though that frequency is typically above Nyquist, and only then aliased in amongst other ones).

Do you have a specific application or case for these considerations?

--pt