Econ Papers Daily / 期刊论文
重新审视倾向得分匹配中的自助法推断：对经济学因果估计的启示
Revisiting bootstrap inference in propensity score matching: Implications for causal estimation in economics
摘要
This study examines when bootstrap resampling improves the accuracy of propensity score matching estimators for causal inference with observational data. Existing research focuses on the validity of bootstrap procedures for matching estimators, providing limited guidance on the conditions under which resampling improves estimators. We develop a heuristic framework highlighting the role of a utilization rate, which measures how broadly control units are used across resampled matches, and evaluate its implications through simulations. The results suggest that estimator performance reflects a tradeoff between matching quality and variance reduction through noise aggregation. When treatment assignment is nearly deterministic and overlap is limited, bootstrapping may reduce accuracy. However, bootstrap resampling can improve estimation when treated and control groups exhibit substantial overlap, allowing information to be combined from a broader set of control observations. Results suggest that data structure, rather than computational intensity, is an important determinant of whether resampling improves estimation.
中文摘要
本研究考察在基于观察数据的因果推断中，自助重采样何时能够提高倾向得分匹配估计量的准确性。现有研究主要关注匹配估计量自助程序的有效性，对于重采样在何种条件下能够改进估计量所提供的指导有限。我们构建了一个启发式框架，突出利用率的作用，该指标衡量控制单元在重采样匹配中被使用的广泛程度，并通过模拟评估其含义。结果表明，估计量表现反映了匹配质量与通过噪声聚合实现方差缩减之间的权衡。当处理分配接近确定性且重叠有限时，自助法可能降低准确性。然而，当处理组与控制组存在显著重叠时，自助重采样能够改进估计，因为此时可以结合更广泛的控制观测信息。结果表明，数据结构的而非计算强度，是重采样能否改进估计的重要决定因素。