I’d like to use 3dRSFC and was wondering if I can censor the data with 3dTProject before computing connectivity. Paul mentioned here (https://afni.nimh.nih.gov/afni/community/board/read.php?1,153171,153191#msg-153191) that censoring shouldn’t be done when using 3dRSFC, because there will be missing points. But what if I choose the ZERO or NTPR modes in 3dTProject which preserve the time points, and then compute connectivity with 3dRSFC? Is this a valid approach?

At present, the best way to go to estimate RSFC parameters like ALFF, fALFF, RSFA, etc. if you are censoring would be to use 3dLombScargle (yes, that’s really the name of a program) with 3dAmpToRSFC after. 3dLombScargle can estimate a power spectrum when the sampling rate isn’t constant (i…e, when FMRI data sets have been censored). 3dAmpToRSFC will take 3dLombScargle’s results and estimate RSFC params. Note that if you have a lot of censoring, then the RSFC parameter estimates become biased. In looking at this in an OHBM poster, it seemed like if >20% of your data is censored, then the biasing is too much to get usable results. FOr data with <10% of time points censored, things should be pretty good.

Note that you would likely not want to bandpass in your FMRI processing if you are going to do this. At the very least, if you do bandpass, then you cannot estimate fALFF or fRSFA, and your range of LFFs for ALFF estimation would have to match or be within that bandpass range.

Thanks for your response and really nice to know about these two functions. However, just for my understanding, you say I can’t combine censoring with 3dRSFC because there are missing time points, but isn’t the whole point of the ZERO/NTRP modes in 3dTProject to preserve the same time points as the original time series? If missing time points is the only reason why 3dRSFC can’t be used with censoring, I don’t understand why the ZERO/NTRP doesn’t solve the problem.

Using ZERO in 3dTproject would not be good enough. The issue is that the typical way of estimating spectral power (or amplitude) distributions is via the Fourier Transform (FT, or Discrete Fourier Series (DFT)). To use the FT, the time series data must be regularly sampled. Inserting zeros means that the length of time series doesn’t change, but those would be interpreted as values of the time series and used in the FT-- that would change the spectrum a lot, for residual time series like resting state likely increasing the relative power of low frequencies artificially.

Using NTRP might be a bit better in some sense, but it would still likely lead to a bias toward low frequencies (if the interpolation is low order) or potentially some artificial high frequencies. It might be hard to predict the effect here, but the main thing is that one is still “tampering” with the final power spectrum, changing its values solely by the censoring processing.

What one would prefer would be to ignore those censored time points (which is different than using them with a value of 0 or of an interpolated value), and estimate the power spectrum of just the time series points that were not censored. The issue there is that once censoring has occured, the remaining time series is no longer regularly/uniformly sampled (because there gaps chopped out of it). Fortunately, astrophysicists were having to struggle with this a while back because their time series of astronomical observations would get interrupted by moving planets, moons, clouds, etc. A number of people worked on this, but the ones who get hte most credit are Lomb and Scargle in the late 1970s/early 1980s, who worked out an approach that was equivalent to classical FT results if the data were uniformly sampled, and then a reasonable extension with similar statistical properties when the data were not uniformly sampled. There has been further work on this still, using further adaptations of their work and other ones (here is an excellent one, if you are interested: https://arxiv.org/abs/1703.09824).

Thank you very much Paul for this very informative answer. Everything you said makes a lot of sense. Are there any plans to include 3dLombScargle and 3dAmpToRSFC in the afni_proc processing pipeline any time soon?
Best regards,
Sam

The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.