Hi, Clément-
FMRI (= EPI) datasets don't have physical units when they are plopped out of the scanner; they are relative to that scan. You could scan me today and tomorrow on the same machine, and the baseline value of BOLD signal could be very different-- this means that any comparison of something that depends on the units of BOLD can't be done directly.
Scaling the BOLD time series produces a dataset that has a meaningful set of units: local (=per voxel) BOLD percent signal change. This *is* something that could reasonably be compared across people at different times (assuming no other large/significant change in the scanner settings/behavior).
When people acquire resting state data and calculate correlation coefficients, the scaling doesn't matter; you can take 2 time series, and either multiply each one by a constant or add a constant (or both)-- it won't affect the correlation values. Therefore, people don't necessarily scale RS-FMRI data.
However, ALFF carries units from the original time series: it is a sum of amplitudes, and those have physical units. So, if you wanted to compare such values across subjects, I don't see how you couldn't *not* scale. Note that as a sum of amplitudes, the length of time series will in general also matter for fair comparisons. So, if the two groups you are comparing have different lengths of time series, that wouldn't be a fair comparison. Also, you are (hopefully) censoring your data, which would effectively alter the length of your time series; *however*, assuming you are using the recommened 3dLombScargle+3dAmpToRSFC with your censored time series to estimate the ALFF and other such parameters, then actually the amplitudes should be reasonably scaled so that the overall spectrum has similar magnitude-- but if there is a lot of censoring, say more than 20% of time points censored, then that comparability breaks down a bit.
fALFF is a dimensionless ratio, so it should be fairly agnostic to scaling/units. Note that again, if you have a lot of censoring, your spectral estimates underlying ALFF and fALFF estimates will start getting biased.
I don't know how often people worry about these things in applications in the literature, but they should...
Can I ask what is your typical range of censored time point fractions?
--pt