Matching Methods in Practice

Causal Question / Estimand

The ATE under unconfoundedness — a hands-on guide translating the theory of Imbens2004-NonparametricATEReview into a recommended workflow, worked through three examples (including the Lalonde/Dehejia–Wahba data).

Identification Strategy

Same selection-on-observables foundation — Ignorability plus Overlap. The contribution is procedural: (1) assess and trim to ensure Overlap; (2) estimate and check the Propensity-Score for covariate balance; (3) estimate the effect with a method that is robust to model misspecification — Imbens recommends combining matching or subclassification on the propensity score with regression adjustment (a doubly-robust, bias-corrected estimator); (4) run supplementary analyses probing the plausibility of unconfoundedness (e.g. pseudo-outcomes, placebo treatments).

Key Assumptions

Ignorability (unconfoundedness), Overlap, SUTVA, and adequate covariate balance after conditioning on the Propensity-Score.

Threats to Validity

Limited overlap (the most common practical failure) — addressed by trimming/discarding units outside common support; reliance on a single estimator without robustness checks; unconfoundedness itself, which is untestable but can be stress-tested.

Setting / Data

n/a — methodological/practitioner guide; illustrated on the Lalonde (1986) NSW data in the Dehejia–Wahba version.

Key Claims

  • Always inspect overlap first and trim before estimating; poor overlap is the leading source of unreliable matching estimates.
  • Combine the propensity score with regression adjustment for robustness rather than relying on matching alone.
  • Although unconfoundedness is untestable, partial assessments (placebo outcomes, estimating “effects” on pre-treatment variables) raise or lower its credibility.

Connections

Citation

Imbens, G. W. (2015). Matching Methods in Practice: Three Examples. The Journal of Human Resources, 50(2), 373–419.