Hi Guiseppe,
Thanks for the note about 3dSeg, we will look into that.
Regarding the bandpass step, that is how bandpassing works. Most people
do not get to see this fact because they just run some FFT as a single
magical step. Whether bandpassing is a good idea or appropriate, I will
leave for others to discuss. But it certainly is common, and this is
the effect.
Since your TR is 2.0 seconds, the Nyquist frequency is 1/4 = 0.25. And
assuming you followed the example and used a max frequency of 0.1, that
means you are effectively giving up ~60% of your degrees of freedom for
the bandpass operation (frequencies from 0.01 or 0.1 are kept, 0.1 to
0.25 are discarded). And 60% of 200 TRs is 120. A few more are lost at
frequencies below 0.01 in the example.
At a slower TR, you would actually give up less, because it is the high
frequencies that are being discarded. With a TR = 1.0s, 80% of the dof
would go to the bandpass step! With a TR = 5s, you would not give up
any dof to high frequencies (because they cannot be modeled, they are
all aliased).
To think of it another way, if one were to run a bandpass operation of
all frequencies up to the Nyquist, the time series would be completely
decomposed (all dof used).
This is the (almost always hidden) cost of bandpassing.
- rick