Hi Judd,
If there were NT time points, then NT basis functions would
be used to fully fit the time series using sines and cosines.
There is no extra doubling. Note that use of sines and
cosines is equivalent to using complex numbers. One does not
need to do this in the complex domain.
The whole point of choosing bandpass basis functions is to
model NT time points up to the Nyquist frequency with NT
terms, not 2*NT. As an aside, it is not 3dDeconvolve doing
this, but 1dBport, the output of which is then passed to
3dDeconvolve as regressors of no interest.
You have 132 time points remaining after censoring, which is
not enough to estimate the 177 parameters from bandpassing,
motion and baseline.
Censoring 50% of your TRs should be a big red flag. Recall
from my first post that you will give up ~60% of your DOF
just for the bandpass regressors, forget motion, etc. So
censoring even 40% is already a deal-breaker.
---
To state the computation another way, if every frequency
over 0.1 is regressed out, then the remaining fraction of
DOF available for regression is 0.1/Nyquist = 0.1*2*TR.
In your case, with TR=2, there will be 0.4 (40%) of your
TRs available for DOF after bandpassing. The same fraction
as noted above. So again, 50% censoring is way too much.
- rick