Randomization

Assigning units to treatment or control by a known chance mechanism. Randomization makes treatment status independent of potential outcomes, so the difference in observed group means is an unbiased estimate of the average Causal-Estimand — it delivers Ignorability by design rather than by assumption.

Fisher established randomization as the basis of valid experimental inference; Rubin’s potential-outcomes formalization (Rubin1978-BayesianCausalEffects, Rubin1980-RandomizationAnalysis) clarifies what randomization buys and what still rests on assumptions (e.g. SUTVA).

Relied on by

RCT (directly); the benchmark against which observational methods (PSM, DiD, RDD, SCM) are judged.

Referenced by

Rubin1974-EstimatingCausalEffects, Rubin1978-BayesianCausalEffects (central even in Bayesian inference), Rubin1980-RandomizationAnalysis, Holland1986-StatisticsAndCausalInference.

Credibility-revolution pass (RCT benchmark / as-good-as-random variation / randomized adoption timing / instrument independence): AngristPischke2010-CredibilityRevolution, AngristKrueger1999-EmpiricalStrategiesLaborEconomics, Angrist2022-EmpiricalStrategiesIlluminatingPath, Meyer1995-NaturalAndQuasiExperiments, AtheyImbens2022-DesignBasedDiD.

Method-buildout pass (2026-06-17) — the RCT literature (AtheyImbens2017-EconometricsOfRandomizedExperiments, BruhnMcKenzie2009-InPursuitOfBalance, DufloEtAl2020-InPraiseOfModeration, LudwigEtAl2011-MechanismExperiments, AngelucciDiMaro2016-SpilloverEffects); the local-randomization framework of RD (CattaneoEtAl2020-RDDHandbook, and as-good-as- random in LeeLemieux2010-RDDInEconomics); instrument independence in AngristKrueger2001-SearchForIdentification. Design-based standard errors: AbadieEtAl2020-SamplingVsDesignUncertainty.