Most evaluations of education policies focus on their mean impacts; when distributional effects are investigated it is usually by comparing mean impacts across demographic subgroups. We argue that such estimates may overlook important treatment effect heterogeneity; in order to appreciate the full extent of a policy's distributional impacts one should also exploit alternative methods. We demonstrate this using data from Project STAR, where we find evidence of substantial treatment effect heterogeneity across achievement quantiles. While all children appear to benefit from being placed in small classes, the largest test score gains are at the top of the achievement distribution. This result seems to be at odds with previous evidence that smaller classes benefit disadvantaged children most, but the discrepancy is reconciled by the fact that there are similar patterns of treatment effect heterogeneity within demographic groups, and that gains for disadvantaged students are larger throughout much of the achievement distribution. (Contains 4 tables and 6 figures.)