Hi Lee,

> Let's further assume that we want to use the beta-weights and
> the intercept estimate to give us percent change, as in
> 100*beta/intercept.
>
> Obviously, this will only work with 1 particular scaling of the regressor.

If you scale the original signal by its mean for each voxel before running individual subject regression analysis, you would get around this problem. You can argue that such scaling is not accurate because mean is usually slightly bigger than intercept, but such inaccuracy is pretty negligible.

> But if I want to get percent change at the peak of the response,
> using 100*beta/intercept, don't I need to have the peak == 1 ?

This whole business of percent signal change is a little dicey to me: % signal change relative to the impulse response function or the whole HRF? This matters with block designs because you would scale the beta's differently.

> If I use waver with -peak 1 will that give me a peak of 1.0, an
> integral of 1.0 or what?

The default peak value (option -peak) for waver is 100. waver scales the impulse response function before the convolution is done. For example, with event-related design, the final peak of HRF is

(duration of the stimulus)/TR * peak

as seen in

waver -GAM -dt 2 -peak 1 -tstim 0:0.4 7.7:8.1 | 1dplot -stdin &
waver -GAM -dt 2 -tstim 0:0.4 7.7:8.1 | 1dplot -stdin &

Similar thing happens in block design:

waver -GAM -TR 1 -peak 1 -inline 16@1 40@0 16@1 40@0 | 1dplot -stdin &
waver -GAM -TR 2 peak 1 -inline 16@1 40@0 16@1 40@0 | 1dplot -stdin &

Well, I feel this probably doesn't matter much in the end. As long as you scale the original signal or beta's consistently across subjects, it should be fine at group level because it is the statistics that people pay more attention to. And the real percent signal change value is not really reliable considering various scaling approaches.

Gang