方法对比

	Natural Experiment in Politics	Difference-in-Means Estimator
领域	Political Science	Political Science
方法族	Process / pipeline	Process / pipeline
起源年份	2012	1923
提出者	Dunning (design-based framework); Lee (close-election RD lineage)	Jerzy Neyman (design-based potential-outcomes framework)
类型	Observational study exploiting as-if random assignment	Design-based estimator of the average treatment effect
开创性文献	Dunning, T. (2012). Natural Experiments in the Social Sciences: A Design-Based Approach. Cambridge: Cambridge University Press. ISBN: 9781107698000	Gerber, A. S., & Green, D. P. (2012). Field Experiments: Design, Analysis, and Interpretation. New York: W. W. Norton. ISBN: 9780393979954
别名	Political natural experiment, As-if random design, Design-based natural experiment, Quasi-experiment with as-if randomization	Neyman estimator, Design-based ATE estimator, Difference of sample means, Mean-difference treatment effect estimator
相关	4	4
摘要	A natural experiment in political science exploits a naturally occurring source of as-if random assignment — close elections, lotteries, arbitrary boundaries, or policy thresholds — to identify causal effects without the researcher manipulating anything. Codified for the social sciences by Thad Dunning's 2012 design-based treatment and exemplified by David Lee's close-election regression-discontinuity analysis of U.S. House races, the approach treats nature, institutions, or chance as if they had run an experiment, recovering credible causal estimates from observational data when randomization is impossible.	The difference-in-means estimator is the design-based workhorse for analyzing randomized experiments: it estimates the average treatment effect simply as the difference between the average outcome among treated units and the average outcome among control units. Rooted in Jerzy Neyman's potential-outcomes framework and central to modern treatments by Imbens and Rubin and by Gerber and Green, it is unbiased under randomization, comes with a conservative Neyman variance estimator, and supports exact randomization inference, requiring no model of how outcomes are generated.