Randomization Analysis of Experimental Data: The Fisher Randomization Test (Comment)
Causal Question / Estimand
In a paired-comparison experiment, the unit-level effect and its average over units — the standard Causal-Estimand — analysed through the lens of Fisher’s randomization test.
Identification Strategy
A comment on Basu that doubles as a foundational clarification. Rubin defends the Fisher Randomization Test as logically viable and frames causal inference as a missing-data problem: each unit’s other potential response is simply missing, and Randomization provides the basis for inference about the unobserved values.
Key Assumptions
SUTVA — named explicitly here for the first time (“stable unit-treatment value assumption”): no interference between units and no hidden versions of treatments. Also Potential-Outcomes and Randomization.
Threats to Validity
SUTVA violations — interference between units (spillovers) or multiple versions of a treatment (“technical errors”) — break the representation on which the analysis rests. Randomization-test inference is conservative but, per Rubin, not “illogical.”
Setting / Data
n/a — theoretical/methodological comment. Paired-comparison experiment as the working example.
Key Claims
- Coins and defines SUTVA, isolating the assumptions that make potential outcomes well-defined.
- Defends Fisher’s randomization test against Basu’s “not logically viable” charge.
- Articulates the missing-data view: because each unit is exposed to only one treatment, the other outcome is missing and must be inferred.
Connections
- Builds on: Rubin1974-EstimatingCausalEffects
- See also: Holland1986-StatisticsAndCausalInference, Rubin1978-BayesianCausalEffects, Foundations
Citation
Rubin, D. B. (1980). Comment on “Randomization Analysis of Experimental Data: The Fisher Randomization Test.” Journal of the American Statistical Association, 75(371), 591–593. https://www.jstor.org/stable/2287653