Hi Mahen,

The mean of baseline-removed signal would typically not be
exactly zero but close to it (it would be the mean of the
signal of interest, which might be 1 or 2 for a very active
voxel, say). In any case, such a scaling would surely not
do anything helpful.

The trend does not really affect the scaling since Legendre
polynomials are used to model the baseline, and they are all
zero-mean except for the constant term.

So the only difference between scaling with the mean and with
the baseline is maybe the 1-2% mean of the non-baseline
regressors (usually of interest). That does not mean a beta
of 1.5 would re-scale to a beta of 3.5, but that it might
re-scale to 1.53, which you would not even notice (especially
since all subjects are scaled the same way).

---

Regarding how well the model fits the data, just plot the
fit time series (fitts from 3dDeconvolve and afni_proc.py).
For details, see page 18 (t18_results_2_EPI.txt) from the
class data tutorial:

afni.nimh.nih.gov/AFNI_data6/tutorial

Alternatively, look at the residual time series (but again,
you would prefer this to be a scaled dataset).

Of course, the full F-stat shows how much variance the non-
baseline parameters account for when compared with only the
baseline model.

Yes, significance in the baseline parameters means they were
useful in modelling the data. But for scaling, only pol#0_Coef
terms are useful.

---

For deciding on the polort, we (Bob) suggest using 1 + the
duration of a run (in seconds) / 150. That is the default
in both 3dDeconvolve and (therefore) afni_proc.py.

- rick