We offer a new analysis of a large-scale trial of an early-childhood education program that targeted premature, low-birthweight children. This targeting heavily oversampled twins, whose outcomes differed significantly from singletons'. Singletons' gains in short-term cognition and age-18 non-cognitive skills were comparable to those of the Perry Preschool and Carolina Abecedarian Projects, supporting those programs' scalability. For twins, however, the program generated smaller positive short-term gains and negative age-18 impacts. These outcome differences arise from differences in parents' response to the program. A household production model suggests that the possibility of jointly supplying parenting to twins helps explain those differences.