Dear all

I recently tried to include a hi-pass filtering scheme in 3dDeconvolve by including in the baseline model a Discrete Cosine Transform basis set, modeling (and therefore regressing out) low-frequency trend. This approach is the one adopted in SPM and it seems to work pretty well. The cosine series is:

t - t0
f (t) = cos(r * pi * ------- )
r tN - t0

where:

r=1,2,...,R
t0,tN=start and end times of the epi sequence

A reasonable default cutoff for R correspond to a lower bound of 1/128 Hz, based on the 1/f spectrum of mri noise and the shape of the HRF.

In my specific case, since I have a very long epi run (~35 min), the number of needed cosines is 32. If we also add the 6 motion parameters and the 2 parameters of a 1st order baseline, we end up with a 40 parameters baseline model.

Now, chances are that there is some degrees of collinearity within this massive baseline model (e.g., between the 1st order baseline and the cosine component with r=1, or between any of the cosine regressors and the motion parameters). Since we are not generally interested in the specific fit of the regressors in the baseline model, but we simply want to 'model out' the confounds, would it be reasonable to reduce the regressors in the baseline model to a smaller orthonormal set and use that in 3dDeconvolve?

thanks for any thought on the subject

giuseppe