标题： 基于设计的随机化、选择和缺失性测试：利用广泛可获取的卫星图像进行远程审计 显示英文标题

标题： Remote Auditing: Design-based Tests of Randomization, Selection, and Missingness with Broadly Accessible Satellite Imagery

摘要： 随机对照试验（RCTs）是因果推断的基准，但实地实施可能偏离注册的设计，或者偶然导致不平衡。 我们引入了一种远程审计——一种可预先注册的设计诊断方法，使用严格治疗前、公开可用的卫星图像来检验分配是否独立于当地条件。 该审计实施了一个条件随机化检验，询问治疗是否比注册机制下更容易从治疗前特征中预测，提供了一个有限样本有效的非参数检查，尊重分块和聚类，并通过最大统计量控制图像模型、分辨率和补丁大小之间的多重性。 相同的预先注册程序可以在基线数据收集之前运行以指导实施，并在分配完成后用于审计实际分配。 在两个示例中——乌干达的青年机会计划（随机化得到证实）和孟加拉国的一项基于学校的实验（相对于设计，分配可预测，与独立担忧一致）——该审计可以在昂贵的科学投资之前及早发现潜在问题。 我们还提供了关于研究参与和缺失性的描述性诊断。 由于其低成本且可以在不同全球行政司法管辖区以统一方式快速实施，远程审计补充了平衡测试，加强了预先注册，并在传统数据收集缓慢、昂贵或不可行时实现了快速设计检查。 显示英文摘要

摘要： Randomized controlled trials (RCTs) are the benchmark for causal inference, yet field implementation can drift from the registered design or, by chance, yield imbalances. We introduce a remote audit -- a preregistrable, design-based diagnostic that uses strictly pre-treatment, publicly available satellite imagery to test whether assignment is independent of local conditions. The audit implements a conditional randomization test that asks whether treatment is more predictable from pre-treatment features than under the registered mechanism, delivering a finite-sample-valid, nonparametric check that honors blocks and clusters and controls multiplicity across image models, resolutions, and patch sizes via a max-statistic. The same preregistered procedure can be run before baseline data collection to guide implementation and, after assignments are realized, to audit the actual allocation. In two illustrations -- Uganda's Youth Opportunities Program (randomization corroborated) and a school-based experiment in Bangladesh (assignment predictable relative to the design, consistent with independent concerns) -- the audit can surface potential problems early, before costly scientific investments. We also provide descriptive diagnostics for selection into the study and for missingness. Because it is low-cost and can be implemented rapidly in a unified way across diverse global administrative jurisdictions, the remote audit complements balance tests, strengthens preregistration, and enables rapid design checks when conventional data collection is slow, expensive, or infeasible.