I have preprocessed a set of resting state functional images through fMRIprep version 20.2.6. For the denoising step, using AFNI’s 3dTproject, I would like to regress out the first five principal components generated from aCompCor, 6 head motion parameters and their derivatives, apply a simultaneous bandpass filter (0.01 - 0.08 Hz), and detrend. I saw on the fMRIprep outputs page that it is recommended to include the cosine regressors in the regression matrix when using Compcor components. If my understanding is correct, by including the cosine regressors, the data will be high pass filtered to retain frequencies above 0.008 Hz. I was wondering if it is appropriate to apply bandpass filtering when the cosine regressors are also included in the regression matrix (i.e., double filtering)?
The cosine filters are a highpass filter. While it is theoretically okay to bandpass afterwards, you still get different edge effects near the bandpass borders depending on the order and kind of filter you use. Therefore, if you want to use effectively impose a bandpass, you can either 1) use a bandpass that has a lower bound of 0, such that the edge effects should not reach the frequency cutoff already imposed, or 2) just do a low pass at the upper frequency. I doubt the bit of power between 0.008 and 0.01 Hz will cause significant changes in your results.
Thank you so much for your quick and very helpful reply.
So if I’m understanding you correctly the cosine just does highpass filtering. So if I want to include a lowpass filter as well (e.g. filter from 0.1 to 0.008Hz) I should add another bandpass from 0 to 0.1Hz to the same regression?
Do you (or anyone else) have a sense if the lowpass filter influences the correlations significantly?
Yes, that should work. If you are using the CONN toolbox, I think you can just apply a given direction filter (e.g. high or low pass, instead of band pass), though I cannot confirm that. You can also use the Inf flag to remove either bound of the bandpass, as defined here.
We high pass to account for low frequency drift from scanner heating effects and very slow physiological noise. Low pass filtering is meant to address faster physiological noise (such as breathing, heart beat). However, given that cardiac events occur about once a second, one would need to have TRs that are around half a second at most (ideally lower) to effectively resolve them (due to the Nyquist theorem). Thus, low-pass filtering is not perfect, but it is easy to implement
I don’t know if there has been a systematic evaluation of the effect of filtering. It seems to be so ubiquitous that I can’t think of many rs-fMRI functional connectivity studies that do not do some kind of bandpass filtering. Happy to hear others’ thoughts.
