Hi Kara,
> what about voxels that have increased BOLD signal to
> Task A, and even larger BOLD signal to task B? For these
> voxels, the % signal change to Task A would be positive in
> run 1 (when Task A is compared to baseline), but the %
> signal change should be negative (though probably small)
> in run 3 because the baseline for this run was calculated
> based on the mean of activity during Task A and Task B.
> The baseline is should be higher for this run, so the same
> BOLD signal value that was above the baseline in run 1
> would fall below the baseline in run 3, and therefore, the %
> signal change would be negative.
Maybe I misunderstand what you mean here, but the baseline and drifting effect for each run are estimated not based on the mean of activity. Instead they are the "best" fit with least square estimation. So ideally all signal should be correctly detected no matter what their relative magnitudes are. Of course negative regression coefficients do occur in the real world of FMRI data analysis from time to time, and some are real while the others are false. But the reason for false negative beta's is not really because of the relative magnitude of BOLD signal to the baseline and drifting effect, and usually it is due to some other issues such as incorrect timing, bad experiment design, etc..
Gang