Regression Discontinuity
Also known as · RDD · regression discontinuity design · regression-discontinuity · polynomial RDD · causal-inference/RDD
Regression discontinuity design (RDD) exploits a hard treatment-assignment rule based on a threshold of some running variable. Units just above and just below the Cutoff are otherwise comparable — they couldn't perfectly choose which side they landed on — yet one group is treated and the other isn't. Any jump in the outcome at the cutoff must therefore be the causal effect of treatment. RDD has the highest internal validity of any non-experimental design.
When to use
Look for RDD whenever treatment is assigned by a quantitative rule: scholarship cutoffs (test scores), means-tested transfers (income), legal-drinking-age effects (age), close-elections (vote share), pension eligibility (date of birth). PS_5 uses the >50% union-vote rule. The estimand is a LOCAL effect (only for units near the cutoff), valid under the Continuity Assumption. Two main variants: Sharp RDD (treatment is deterministic at the cutoff) and Fuzzy RDD (treatment probability jumps but isn't 0→1, estimated via IV with the cutoff dummy as instrument).