There is a burgeoning body of evidence on the average impacts of educational programs. Yet, for many local decision makers, because impacts can vary across sites, the question of whether a certain program will work in their particular district or school remains. This article addresses the question of the generalizability of large-scale average causal effects to specific locales, that is, from “large to small.” Our method evaluates whether impacts measured on the large-scale generalize to individual sites and whether adjusting for variation among sites in their characteristics accounts for impact heterogeneity that limits generalization. Our application to a multisite experimental evaluation in education shows that the average impact does not readily generalize to sites, with discrepancies between 0.17 and 0.38 standard deviations, and that characteristics of sites account only minimally for this variation. This emphasizes need for caution by local decision makers in accepting average impact findings as generalizable to their specific locales.

关于教育计划的平均影响的证据越来越多。然而，对于许多地方决策者而言，由于影响可能因地点而异，因此某个计划是否会在其特定地区或学校发挥作用的问题仍然存在。本文解决了大规模平均因果效应对特定区域（即从“大到小”）的普遍性问题。我们的方法评估了大规模测量的影响是否可以泛化到单个站点，以及调整站点特征之间的差异是否会导致限制泛化的影响异质性。我们在教育中的多站点实验评估中的应用表明，平均影响不容易推广到站点，标准差在 0.17 和 0.38 之间，并且网站的特征只是最低限度地解释了这种变化。这强调了当地决策者在接受可推广到其特定地区的平均影响调查结果时需要谨慎。