Representation Matters: Assessing the Importance of Subgroup Allocations in Training Data