Hi Jinny-
Voxels that are "positive" are postively correlated with the predicted response while those that are "negative" are anticorrelated with the predicted response. If some regions of the brain consistently show negative activity, this might indicate a functional distinction from those regions that consistently show positive activity. The idea behind clustering to control for multiple comparisons is that voxels in spatial proximity respond similarly. If you maintain this assumption of functional similarity of spatially proximal voxels, you should cluster positive and negative separately. This means if you are interested in negative activity, you might want to run a separate clustering on the negative voxels. You should get different results when you include or exclude negative voxels. Hope this helps.
Anthony