Randomized Controlled Trials MOC
The randomized controlled trial is the benchmark design of the credibility revolution: physical Randomization makes treatment independent of Potential-Outcomes, so a difference in means identifies the ATE without modelling assumptions. The papers here move from the econometrics of analysing experiments, through the design choices (balance, stratification, Statistical-Power) that determine whether an RCT is informative, to the research-integrity practices (pre-analysis plans, transparency) that keep a body of experimental evidence credible.
Papers
- AtheyImbens2017-EconometricsOfRandomizedExperiments — Fisher (randomization) vs. Neyman (repeated-sampling) inference, covariate adjustment as variance reduction, stratified/paired designs.
- BruhnMcKenzie2009-InPursuitOfBalance — how to randomize in small samples: stratification, pairwise matching, re-randomization, and accounting for the design in analysis.
- Lenth2001-SampleSizeDetermination — credible power/sample-size calculation; against retrospective power.
- DufloEtAl2020-InPraiseOfModeration — how tightly pre-analysis plans should bind: confirmatory vs. exploratory analysis.
- ChristensenMiguel2018-TransparencyReproducibility — publication bias, p-hacking, low power, and the reform agenda (registration, PAPs, replication, open data).
- GerberGreen2012-Attrition — when missing outcomes do and don’t bias an experiment; reweighting under (conditional) MIPO, worst-case bounds, and double sampling.
Key Concepts
Randomization · Potential-Outcomes · Causal-Estimand (ATE) · SUTVA · Statistical-Power · Pre-Analysis-Plan · Attrition · LATE (noncompliance)
Debates & Contradictions
- How rigid should pre-analysis plans be? ChristensenMiguel2018-TransparencyReproducibility champions PAPs as a credibility fix; DufloEtAl2020-InPraiseOfModeration cautions that over-specified plans suppress legitimate learning — “moderation.”
- Fisher vs. Neyman inference. Sharp-null randomization tests and repeated-sampling ATE inference answer different questions; both are design-based (AtheyImbens2017-EconometricsOfRandomizedExperiments).
- Pure randomization vs. balancing. Stratification/matching improve finite-sample balance but must be reflected in analysis (BruhnMcKenzie2009-InPursuitOfBalance).
- Point estimates vs. bounds under attrition. When missingness may depend on potential outcomes, reweighting needs an (untestable) ignorability assumption, while worst-case bounds give an assumption-light but wider answer (GerberGreen2012-Attrition, Attrition).
Next
The RCT is the design all observational strategies emulate — Foundations, Randomization. Noncompliance turns an RCT into an IV problem (IV, LATE); the local-randomization view of RD (RDD) borrows its logic.