Temporal Filtering Papers

Useful Papers - Temporal Filtering


Friston et. al (2000), "To smooth or not to smooth?: bias and efficiency in fMRI time-series analysis," NeuroImage 12, 196-208 PDF

Summary: The major last defense of low-pass filtering (aka temporal smoothing) mounted by the SPM group before abandoning it for a different mathematical tack in estimation, more amenable to pre-whitening. The case made is that pre-whitening, while more efficient at removing variance from the timecourses, can be more sensitive to bias due to errors in estimating the autocorrelation, and so band-pass filtering has its place.

Bottom line: Improvements in autocorrelation estimation and robustness of pre-whitening have rendered this paper basically obsolete, but it's a good look at the very detailed issues surrounding the apparently simple question of temporal filtering.

Purdon & Weiskoff (1998), "Effect of temporal autocorrelation due to physiological noise and stimulus paradigm on voxel-level false-positive rates in fMRI," Human Brain Mapping 6, 239-249 PDF

Summary: Clear and relatively concise look at exactly how noise autocorrelation can bias inferences in both directions, with specific references to existing models like the extended GLM of Friston et. al and the AR(1) model of Bullmore et. al. Proposes an AR + white noise model to model temporal autocorrelation (which ends up being largely adopted in SPM2)

Bottom line: One of the clearest and most intuitive overviews of what the issue is with temporal autocorrelation and modeling it.


Skudlarski et. al (1999), "ROC analysis of statistical methods used in functional MRI: individual subjects," NeuroImage 9, 311-329 PDF

Summary: Once again, Skudlarski et. al's look at single-subject simulated data, analyzing various preprocessing choices with receiver operating characteristic (ROC) curves, measuring both types of error at a given threshold. They show convincingly that low-pass filtering with real fMRI noise hurts sensitivity, while high-pass filtering can help - although only a precise cutoff threshold choice improved high-pass filtering above simple quadratic and linear trend removal.

Bottom line: Linear and quadratic trend removals are a must. Above that, if you can get in the ballpark of a good cutoff threshold, a full high-pass filter is good, too. Low-pass filtering's nothing but trouble.

Della-Maggiore et. al (2002), "An empirical comparison of SPM preprocessing parameters to the analysis of fMRI data," NeuroImage 17, 19-28 PDF

Summary: Another look at this paper, which examines various preprocessing choices from simulated data with the twin measures of power and false-positive rate. They find high-pass filtering to be useful with a fixed ISI (but not particularly for variable ISIs). They find both low-pass filtering and pre-whitening to decrease power but only pre-whitening to be especially good at protecting against false-positive increases.

Bottom line: High-pass is good for fixed-ISI studies but maybe useless in variable-ISI studies; low-pass filtering tends to decrease power without protecting against inferential bias particularly well.

Holmes et. al (1997), "Statistical modelling of low-frequency confounds in fMRI", NeuroImage 5, S480 (PDF not available yet - see Jeff for paper copies)

Summary: An oldie but a goodie, Holmes et. al shows the first published example of a design matrix including the high-pass filter that becomes a staple of later versions of SPM. The confounds are modeled in basic but effective fashion - a set of cosines of gradually increasing frequency, up to a certain cutoff point. They demonstrate that these regressors account for a good chunk of noise variance.

Bottom line: High-pass filtering works to absorb noise; modeling with cosines is effective and computationally tractable.