I agree with Brian and want to add that it is important that you model as much of the variance as you can, and if you model A and B together you might have less flexibility to model the variance and therefore the goodness of fit of the model in some voxels might be significantly affected (i.e. reduced). So even if you are interested in C, the voxels that show some effect to C might also be activated or deactivated to A or B. If they are not activated or deactivated to A and B precisely the same then your explained variance will be reduced (i.e. your residuals == error will be greater) and therefore you might miss the effect of C that you are looking for (since such voxels might not come out as significant; significance is affected by the error term).

Good luck

Sharon