Philippe,

(1) By converting the original signal with (x/Mean) * 100, you basically scale the data and end up with numbers around 100. If you make your regressors with a magnitude of 1 by setting waver with option -peak 1, each regressor coefficient would automatically reflect the percent signal change for that condition/task/stimulus since the regression model takes care of the average with the baseline component.

(2) Do you model the two runs separately in regression analysis so that you could compare the percent signal change between them? If so, scaling with (x/Mean) * 100 (here Mean would be obtained in separate procedure for each run) would be good enough for such a purpose.

(3) By (x – Mean / Mean ) * 100), I suppose that you really meant ((x – Mean)/Mean)*100). Otherwise you were in fact doing (x-1)*100. With ((x – Mean)/Mean)*100), other than scaling, you also shifted the signal around its mean, which should not cause any effect on the final results of regressor coefficients other than slighly different baseline modeling. And you might not need to run 3dDetrend although it would not do any harm either. I have problem understanding why you would get extremely high values of percent signal change inside of the brain.

(4) 3dAutomask would basically employ 3dClipLevel and a few other procedures to generate a masking dataset. It is optional and only for cosmetic purpose.

Gang