Dear Hikaru
It seems, that there are quite a few approaches for preprocessing of task-free fMRI data, but from the conceptual point of view it may be important to differentiate between noise-modelling and band-pass filtering to frequencies between ~0.001 - 0.01 Hz.
Whereas low-pass filtering the data may (will) reduce higher frequency noise, explicitly modelling different components of noise and/or systematic signal variations in the time-series data should be more effective in reducing the effect of nuissance varaibles. Without claiming to be exhaustive, common appraoches include realignment-parameters (entering as first-order or first- and second-order terms), the derivative of the movement parameters (movement from the previous scan to the present, which again may be entered as first- or first/second- order terms), mean GM/WM/CSF-signal as obtained from, e.g., tissue-probability masks or from voxel placed at representative locations into these tissues, and finally physiological noise (respiration, pulse) correction approaches for which again various forms exist. While there seems no absolute agreement on which of these components remove imaging-related noice without introducing systematic confounds (cf. the anti-correlations debate), the key point here is, that noise introduced by known confounds should be modelled and removed.
There is, however, unrelated to the above, a second, physiological, argument for low-pass filtering: Evidently it is assumed, that meaningful MRI signal changes and hence functional connectivit as identified by their correlation, should be based on neuronal activity changes (though cf. the consideration by Raichle and colleagues on a potentially metabolic basis). Evidently, however, any neuronal activity should be expressed in fMRI only by the hemodynamic response function, which effectively actis as a low-pass filter. Hence, any signal that is to high-frequent to reflect neuronal signal convolved with the hrf should be (e.g., unsystematic or measurement) noise.
In other words: The "standard" preprocessing combining noise modelling and low-pass filtering is based on two considerations, the removal of known confounds and the focus on those signal-frequencies that could reflect hrf-convolved neuronal activity.
Not sure, that AR(1) can solve both of these problems.
Best regards
Simon
Hi SPM users of resting state fMRI researchers,
I have a question regarding the use of autoregressive model over low-pass fileter in resting state fMRI analyses.
I analyzed the data of rsfMRI by investigating the voxels in the whole-brain, which signal correlate with those of ROIs.
In the analysis, I have used the auto-regressive model instead of low-pass filter. Serial correlations in the BOLD signal were accounted for by a first-degree autoregressive correction (Della-Maggiore et al. 2002) (instead of being removed by a low-pass filter).
But it seems multiple reviewers had problems in use of autoregressive model, intead of low pass filter in rsfMRI analyses. They say low-pass filter has been widely used and recommend low-pass filter.
The use of autoregressive model over the use of low-pass filter has been recommended when the signals of the model frequently changes through time series such as in rapid presentation event related designs (Della-Maggiore et al. 2002). But it seems the reviewers are not pursuaded by this reference and it is true this 2002's paper did not tell anything regarding the use of autoregressive model in resting state fMRI analyses, it just talks about the cases where the signals of the model frequently changes through time series.
From what I understand, the low-pass filter has no superiority over autoregressive model in normal fMRI analyses and hence this option is explicitly removed from SPM after a debate long time ago and now no longer available easily. Are there any particular theoretical reasons or evidences that low-pass filter should be used instead of autoregressive model or any reasons or evidences against the use of low-pass filter in rsfMRI analyses (over autoregressive model)? Or is low-pass filter used over autoregressive model just because everyone is using it?
I frequently recieve similar critiques regarding this point in different studies but I am not good at the theoretical background of these things and I hope someone can help me.
Hikaru