Hi all,
I’m a postdoctoral researcher working on dynamic functional connectivity analysis (DCC-GARCH) applied to the ADHD-200 dataset, and I’ve hit a serial-correlation problem I’d appreciate the community’s input on.
Context: The BOLD data I’m using was preprocessed with an ideal (FFT-mask, hard-cutoff) bandpass filter at [0.009, 0.08] Hz — no smooth roll-off (AFNI’s 3dBandpass, though the question generalizes to FSL pipelines using comparable hard-cutoff filtering). I’ve confirmed (analytically, via Wiener-Khinchin, and empirically on both real and simulated data) that this guarantees residual autocorrelation at all lags: a rectangular power spectral density implies a sinc-shaped autocorrelation function that never hits exactly zero. AR pre-whitening (orders up to 15-40) and even a Kalman-filter innovations approach both fail to adequately whiten it — Ljung-Box rejects in essentially 100% of series tested.
I recently read Yue, Webster, Grabowski, Shojaie & Jahanian’s “Iterative Data-adaptive Autoregressive (IDAR) whitening procedure for long and short TR fMRI” (Frontiers in Neuroscience, 2024) and its supplement, which report a dramatic improvement on short-TR data — full iterative IDAR bringing whitening failure down from ~100% (raw/AR(1)/ARMA(1,1)) to ~0.3%, at voxel level, whole brain. Very promising — but their test data used a high-pass + narrow band-stop notch filter, a much wider effective passband than my hard-cutoff case, so I’m not sure whether a high-order iterative AR approach can actually overcome a true ideal filter’s mathematically-guaranteed non-whiteness, or whether my case falls outside the conditions IDAR was validated under.
Questions:
- Code availability: Is there a public/existing implementation of IDAR (R or otherwise)? I couldn’t find a repository; Algorithm 1 in the paper is detailed enough to reimplement, but I’d rather not duplicate effort if something already exists.
- Applicability to ideal/hard-cutoff filters: Has anyone tested, or has intuition about, whether IDAR (or a similar high-order iterative AR approach) can adequately whiten data that went through a true ideal/FFT-mask bandpass filter, as opposed to a smoother IIR-type filter (e.g., FSL’s fslmaths -bptf)?
- Implementation language: Was the original IDAR implementation developed in R (e.g., built around forecast::auto.arima)? Just want to confirm before I start reimplementing.
Any pointers, related papers, or a “here’s why this won’t/will work” would be very helpful. Thanks in advance!
Best,
Claudio Avena
Department of Statistics
Federal University of Bahia, Brazil