Hi Long,
It is highly recommended to remove all of the modeled parameters
at once (baseline, motion, stim classes), not piece by piece.
Use 3dDeconvolve to do this.
3dDetrend will remove as much of the motion parameters as it can
to make the best least squares fit to your data. In your case,
since 99% of your data has been subtracted out by that operation,
you are essentially claiming that 99% of your signal is due to
the motion parameters, which is of course, not the case. Did you
remove the polynomial baseline there, too?
It might be much better to let 3dDeconvolve decide how much of
your data is baseline, motion and stimulus response, using just a
single least squares solution.
---
It should not be necessary to orthogonalize the motion parameters,
since you probably do not care to evaluate the results of those
regressors. More likely, you just want them modeled and removed.
- rick