Hi Zhuang,
I did not compute 177 regressors, Judd reported it (from 3dDeconvolve).
But I could certainly approximate it from the:
- bandpass regressors (most of the regressors)
- motion and derivatives
- polort
For bandpassing, the Nyquist frequency is 0.5/TR. We might throw out
all frequencies above 0.1 (faster than 10 second cycle time), so the
fraction of DOF (starting with NT) that are kept is 0.1/Nyquist = 0.2*TR,
while we throw away (regress out) 1-0.2*TR of our DOF.
So suppose TR=2 and NT= 200. Then there are 200 DOF available for a
solvable regression model. Using bandpass regressors to model any
frequency above 0.1 would be done using approximately 1 - 0.2*TR or
.6 of the available DOF, which is to say .6*200 = 120 regressors. So
one would expect to use approximately:
120(bandpass)+12(motion)+4(polort) = 136
regressors in the model.
I don't know TR or NT for the 177 case. That is probably when using a
different TR or NT.
Basically, people using bandpass methods have always had this issue
with loss of DOF. Doing BP via regression just makes it more obvious.
- rick