Identification and Inference in Nonlinear Difference-in-Differences Models

Causal Question / Estimand

The full counterfactual distribution of the treated group’s outcomes absent treatment (and of the control group under treatment), not just a mean. This yields the average treatment effect plus distributional estimands — quantile treatment effects and mean–variance trade-offs (Causal-Estimand).

Identification Strategy

The changes-in-changes (CIC) model — a nonparametric generalization of DiD. Rather than assuming additive parallel trends (which is not invariant to how the outcome is scaled), CIC assumes the outcome is a monotone function of an unobserved unit-level factor whose distribution is stable within a group over time. The control group’s change in the distribution of outcomes identifies the mapping used to construct the treated group’s counterfactual distribution. Reduces to standard DiD under additional restrictions.

Key Assumptions

A scalar unobservable with strict monotonicity of the outcome production function, time-invariant within-group distribution of the unobservable, and Overlap of the support of outcomes across groups (so the control distribution covers the treated). Generalizes — rather than assumes — Parallel-Trends; SUTVA throughout.

Threats to Validity

Failure of the support/overlap condition (control outcomes must span treated outcomes); violation of the within-group time-invariance of the latent distribution; and monotonicity failures. The model trades the scale-dependence of additive DiD for these structural assumptions.

Setting / Data

n/a — econometric methodology. Applicable to either repeated cross-sections or panels; illustrated with empirical examples (e.g. health-insurance and disability-program outcomes).

Key Claims

• Additive DiD is not invariant to monotone rescaling of the outcome; CIC fixes this by modeling the outcome distribution.
• CIC nonparametrically identifies the entire counterfactual outcome distribution, delivering quantile as well as mean treatment effects.
• Standard DiD emerges as a special case of CIC under extra assumptions.

Connections

• Generalizes: the additive Parallel-Trends DiD of AshenfelterCard1985-LongitudinalEarnings and Abadie2005-SemiparametricDiD
• See also: AtheyImbens2022-DesignBasedDiD (the same authors’ design-based take), DiD

Citation

Athey, S., & Imbens, G. W. (2006). Identification and Inference in Nonlinear Difference-in-Differences Models. Econometrica, 74(2), 431–497. https://doi.org/10.1111/j.1468-0262.2006.00668.x