Improving Subgroup Robustness via Data Selection