ALFF fALFF using AFNI

AFNI Message Board
1

AFNI version info (afni -ver):
AFNI_26.1.01, linux_openmp_64, May 26 2026
System: Ubuntu 24.04 Noble Numbat, x86_64

Dear AFNI experts,

I am computing ALFF and fALFF for a resting-state fMRI dataset of Parkinson's disease patients using 3dLombScargle + 3dAmpToRSFC (to handle censored time points as recommended).

My preprocessing used framewise displacement > 0.3mm and outlier fraction > 5% as censoring thresholds (via afni_proc.py). After censoring, some subjects have substantially fewer retained volumes than others, which is expected given the elevated motion in a PD population.

My question is: is there a recommended minimum percentage of volumes (TRs) that should be retained after censoring for ALFF/fALFF estimates to be considered reliable? I haven't been able to find a specific threshold for spectral measures like ALFF/fALFF.

For context, my subjects have TRs of 1.88s, 1.94s, 2.0s, and 2.2s, with total scan lengths of 179 or 249 volumes before censoring. I am currently using 70% retained volumes as a subject-level exclusion threshold, but I would like to know if there is a more principled or literature-supported cutoff I should be using.

Any guidance or relevant citations would be greatly appreciated.

Thank you!

2

Howdy-

Using 3dLombScargle + 3dAmpToRSFC sounds good for that censored data. And that is an excellent question about when censoring becomes "too much" to apply even that, due to biases from losing out a lot of data. (And it is important to note that the L-S method assumes that the censoring is random or quasi-random---periodic censoring could break assumptions in it.)

Here is a poster about 3dLombScargle from ISMRM in 2018 that looked a bit at that question, using repetitions of censoring at various fractions to build up estimation distributions around known values.

  • In that poster, EPI time series distributions with decreasing power spectra (that is, higher magnitude at lower frequencies) are typically going to be the most relevant for GM---that means the "B" and "D" plots there.
  • The maximum of the frequency band is 0.25 Hz (that max is also referred to as the Nyquist frequency), corresponding to a sampling rate, TR=2s.
  • There were 200 time points per time series (and so a total of 400s of data).

As you can see, there isn't a strict cut-off. However, I would say that beyond about 10% of time points being censored (that is, 20 time points out of 200), the distributions aren't so well centered around the actual value, and therefore seem to be more biased.

--pt