Hi Helmut,

Figure 4 in this paper might also be interesting:

The paper suggest that modeling polynomians is more accurate than high-pass filtering. However, they are not explicit about the type of filter they use and I don't see why there should be a huge difference between e.g. DCT filtering vs. adding DCT regressors to the model (which used to be the default in SPM).


2015-06-17 19:21 GMT+02:00 H. Nebl <[log in to unmask]>:
Dear Torben,

Yes, I've checked the implementation in AFNI (Legendre polynomials), the order is based on 1 + floor((duration of the session in s)/150 s), with the 150 s probably based on the same early event-related fMRI studies that resulted in the default 128 s in SPM. Thus in case of block designs with long intervals, long block lengths one would run into the same issues (adjust to HPF to keep slow frequency signal within the model, remove the noise but also remove possibly major parts of the signal) with polynomial regressors. I hadn't really thought about that, as the linked papers dealing with polynomials usually didn't go beyond 4th or 5th order.