The Econometrics of Randomized Experiments
Causal Question / Estimand
The average treatment effect in a randomized experiment, and the inferential frameworks for estimating it and quantifying uncertainty.
Identification Strategy
Identification is by design: physical Randomization makes treatment independent of Potential-Outcomes, so a simple difference in means is unbiased for the ATE — no modelling assumptions required. The chapter’s contribution is to lay out the inference frameworks: Fisher’s exact / randomization tests of the sharp null (treating potential outcomes as fixed, the assignment as the only source of randomness), Neyman’s repeated-sampling approach to unbiased ATE estimation and conservative variance, and the role of covariate adjustment (regression as a variance-reduction device, not an identification device). Extends to stratified and paired randomization, clustered designs, and the analysis of experiments with noncompliance (linking to LATE).
Key Assumptions
Randomization (random assignment), SUTVA (no interference, no hidden versions), and — for compliance/IV interpretations — Exclusion-Restriction and Monotonicity. No functional-form assumptions are needed for identification.
Threats to Validity
n/a for identification under correct randomization; in practice, noncompliance, attrition, interference/spillovers, and small samples (where asymptotic approximations fail and randomization inference is preferable). Covariate adjustment can introduce finite-sample bias if specified post hoc.
Setting / Data
n/a — methodological handbook chapter.
Key Claims
- Randomization identifies the ATE without modelling; regression covariates only improve precision.
- Fisherian (sharp-null, finite-sample) and Neyman (repeated-sampling) inference answer different questions; both are design-based.
- Stratification/pairing and pre-specified covariate adjustment improve efficiency; design choices should be made before seeing outcomes.
Connections
- Builds on: Fisher’s randomization inference (book deferred), Randomization, Potential-Outcomes, Rubin1974-EstimatingCausalEffects
- Design-based inference parallels AtheyImbens2022-DesignBasedDiD and AbadieEtAl2020-SamplingVsDesignUncertainty.
- Noncompliance links to IV/LATE (IV). See also RCT
Citation
Athey, S., & Imbens, G. W. (2017). The Econometrics of Randomized Experiments. In Handbook of Field Experiments (Vol. 1, pp. 73–140). Elsevier.