Hi Paul,

I am not interested in one specific frequency/oscillation; I am not interested in one particular periodicity of the time-series that would then "pop out" to be easily detected in the frequency-domain. Instead, I am interested in the whole frequency spectrum, that is, how the power is distributed, say from 0.01 Hz to the Nyquist frequency.

Now let's get specific (maybe I should have been specific right from the beginning, but I thought it was easier to explain the topic more generally):

I have very long sleep runs up to 3000 time points with a TR of 2.16. The sleep onset heavily diverges between the single subjects. Since the sleep recordings are very long, preprocessing took days. I later realized that I should have cut out the first ~50-400 TRs where subjects were awake, because I am only interested in how the power spectrum looks in sleep. Since the power shifts to slower frequencies in sleep (compared to awake), the power spectra awake vs. sleep look different.

Because I wanted to proceed methodologically correct, I didn't want to cut the initial "awake" TRs of the preprocessed .errts file out now, without further asking you guys if that could lead to a power spectrum that would look different compared to a time-series that was already cut before running AFNI proc on it.
Of course, the power spectrum is not computed by AFNI proc, but since there are so many computations going on in AFNI proc, I was not so sure that one of thos (or the combination and interactive effects of many of those) can subsequently affect the results of 3dPeriodogram.

The bandpassing was 0.01 - 1 Hz. The polort option was set to 2. After running 3dPeriodogram, I cut the power-spectra so that only the frequency range 0.01 - 0.225 remains.

But I already get the feeling it is more safe to cut the data with 3dTcat before preprocessing, and running AFNI proc again on those. Looking forward to your answer/ideas.

Philipp



Edited 3 time(s). Last edit at 05/16/2022 08:58AM by Philipp.