Dear Andy,
I'd say the exact value of 1/128 Hz has no deeper meaning, except that it is 2^7, although it should be a good approximation in general. It traces back to some old event-related fMRI study as far as I know. With standard event-related designs and shorter block designs the expected time course consists of higher frequencies to a major extent, so removing low frequencies (including low-frequency scanner drifts) is a good idea. If the expected time course has peaks in the lower frequency domains (due to long blocks, low stimulus frequency), then the standard high-pass filter might well remove the scanner drifts, but it can also be expected to remove signal. Removal of low frequencies might also introduce cross-correlations between different regressors (different conditions), see https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=SPM;475988d3.1304 Exp1.gif to get an impression of the issue. I rather doubt the "results" are meaningful then. When checking literature you will find papers for each of the two options. Some use the 128 default setting, although one can assume the expected time course to look quite weird, others report a less aggressive filter. Unfortunately many papers don't report the high-pass filter at all.
In one of the few methods papers Skudlarski et al. (1999) conclude: "The best overall result is obtained by use of high pass Butter-worth filtering at the frequency of 0.35 of the stimulus frequency." But we can assume this is related to the specific setting. And there is surprisingly few literature on high-pass filters in the context of long blocks. It would also be interesting to learn how much scanners improved over the last decade.
Concerning increase of false positives with a less aggressive high-pass filer, I'm not sure whether this is true. Smith et al. (2007, Neuroimage) actually conclude the opposite:
"The number of false positives increases with filter cut-off up to 0.022 Hz. The problem is larger by more than an order of magnitude when such a filter is used, compared to when no filter is used. This is true both without [...] and with [...] AR(1). We explain this marked effect of filter cut-off as follows. As the filter cut-off is progressively raised, the analysis becomes progressively more sensitive because the power removed from the timecourses is largely noise, so signal-to-noise improves. In an experiment with a task, this is reflected in enhanced sensitivity to task-related signal at the model frequency. But it also increases sensitivity to noise at the same frequency, so increased sensitivity to signal is accompanied by a commensurate increase in the number of false positives."
This might be related to their strategy though (resting state fMRI was analysed as if there were two blocked conditions).
Best,
Helmut
Best,
Helmut
|