Semiparametric Inference for Functional Difference-in-Differences

Difference-in-differences (DiD) is not a routine scalar generalization; rather, it entails three fundamental challenges in identification, inference, and observation. This paper develops a comprehensive semiparametric inference framework for functional DiD with discretely observed data.

The Core Framework

The authors develop a semiparametric inference framework for causal effects on functional outcomes in a two-period observational DiD design. This framework addresses the challenges of identification, inference, and observation when extending DiD to functional data. The key contributions include formulating and identifying the functional average treatment effect on the treated (fATT) as a curve-valued causal estimand, deriving the efficient influence function for fATT, constructing an associated Neyman-orthogonal, doubly robust score, and incorporating reconstruction from noisy grid observations.

Identification and Influence Function

Identification of the fATT curve is achieved by imposing a functional parallel trends condition conditional on covariates X. The authors derive the efficient influence function (EIF) for the fATT curve, which is expressed in a Hilbert-space formulation. This EIF is shown to be the canonical gradient, or efficient influence function, for the fATT functional defined by equation (5). The second property of this EIF is that the map from nuisance functions to the score is Neyman-orthogonal at η0, which underpins robustness to nuisance estimation error.

Estimation and Inference

The authors propose a cross-fitted doubly robust estimator that utilizes an uncentered orthogonal score. This estimator is implemented using K-fold cross-fitting to mitigate overfitting bias in nuisance estimation. The theoretical results establish asymptotic linearity in a Hilbert space and weak convergence to a Gaussian element in H = L 2(T). This weak convergence allows for pointwise confidence intervals and multiplier-bootstrap simultaneous confidence bands.

Simulation and Application

The paper validates the proposed method through extensive simulations under dense and sparse functional observation designs. Simulation results confirm that the CF–DR estimator achieves superior coverage and testing power in finite samples compared to other methods. The application to London ULEZ and hourly NO2 profiles demonstrates that the method can reveal a time-of-day pattern associated with lower traffic-site NO2, providing a functional causal inference tool.

Conclusion

The proposed CF–DR estimator is shown to be robust and inference-oriented, remaining close to the best feasible estimator in simple settings and becoming clearly stronger in complex designs with flexible nuisance learning. The method successfully separates causal-estimation difficulty from reconstruction-driven functional error. The results support the use of treating the outcome as a curve rather than a scalar average for policy analysis. The paper suggests extensions to staggered adoption and spatial dependence in future work. The theory highlights that curve reconstruction must be accurate enough for first-order causal inference. The paper concludes that CF–DR offers a substantial practical advantage for functional DiD. The authors report that the negative traffic-site pattern persists across various post-policy windows, suggesting a plausible environmental effect. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The simulation results show that increasing sample size improves sampling variability, but reconstruction-induced smoothing can remain visible when the observation process is sparse and noisy. The overall conclusion is that CF–DR remains close to the best feasible estimator in simple settings and becomes clearly stronger in complex dense-observation settings. The paper provides a theoretically grounded and computationally tractable foundation for causal evaluation with functional data. The authors recommend extending the framework to staggered adoption and continuous exposure intensity in future work. The paper highlights that curve reconstruction must be accurate enough for first-order causal inference. The paper concludes that CF–DR offers a substantial practical advantage for functional DiD. The results support the use of treating the outcome as a curve rather than a scalar average for policy analysis. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives. The paper demonstrates the utility of functional causal inference in revealing time-of-day patterns associated with lower traffic-site NO2. The final results indicate that CF–DR provides more balanced uncertainty quantification than main feasible alternatives.

1. Improved Causal Inference for Functional Outcomes: The system can now perform causal inference on continuous data such as circadian activity curves or long-term biomarker trajectories by estimating the functional average treatment effect on the treated (fATT) curve tau0(t).

2. Enhanced Robustness to Nuisance Misspecification: By employing a cross-fitted doubly robust estimator, the AI system can maintain first-order validity even when nuisance functions like propensity scores or outcome regressions are misspecified, as shown by its performance across scenarios S3 and S5.

3. Precise Curve-Level Uncertainty Quantification: The system can generate asymptotically valid simultaneous confidence bands for the entire functional domain T, moving beyond scalar summaries to quantify uncertainty over the whole trajectory using a multiplier bootstrap approach.

4. Handling Sparse and Noisy Measurements: The improved framework incorporates reconstruction from noisy grid observations, allowing the AI to produce causal estimates even when functional outcomes are observed on sparse or irregularly sampled data, provided reconstruction error is asymptotically negligible.