Semiparametric Difference-in-Differences Estimators

Causal Question / Estimand

The average treatment effect on the treated (ATT, Causal-Estimand) when the standard DiD parallel-trends assumption is implausible unconditionally because treated and control groups differ in covariates that drive outcome dynamics.

Identification Strategy

Relaxes unconditional Parallel-Trends to Conditional-Parallel-Trends: parallel paths hold only after conditioning on observed covariates . Identification then proceeds by reweighting — an inverse-propensity-score (IPW) two-step estimator that reweights the control group’s before/after change to match the treated group’s covariate distribution. Also lets the ATT vary with , describing effect heterogeneity across observed characteristics.

Key Assumptions

Conditional-Parallel-Trends (parallel trends given ), Overlap (propensity score bounded away from 0 and 1 so controls exist for every treated covariate profile), and SUTVA.

Threats to Validity

Misspecification or omission of the covariates that govern outcome dynamics (e.g. the Ashenfelter-Dip driven by transitory pre-treatment shocks); propensity scores near the boundary that break Overlap and inflate variance.

Setting / Data

n/a — methodological/econometric. Two-period treated-vs-control panel or repeated cross-section; illustrated with a JTPA training application.

Key Claims

  • When covariates create non-parallel raw trends, conditioning restores identification — unconditional DiD is biased but conditional DiD is not.
  • A simple propensity-score-weighted two-step estimator recovers the ATT semiparametrically, without a fully specified outcome model.
  • The framework characterizes how the treatment effect varies with observed characteristics.

Connections

Citation

Abadie, A. (2005). Semiparametric Difference-in-Differences Estimators. The Review of Economic Studies, 72(1), 1–19. https://doi.org/10.1111/0034-6527.00321