Debiased Semiparametric Efficient Changes-in-Changes Estimation

2025-07-09

AI Summary

This paper addresses the challenge of identifying average treatment effects on the treated (ATT) and distributional causal effects in panel data with multidimensional unobserved confounders. To overcome the limitations of conventional Change-in-Changes (CiC), which relies on strong assumptions of a single latent variable and monotonicity, we propose a generalized CiC framework that permits multidimensional, non-monotonic confounding. We construct a semiparametric estimator with Neyman orthogonality, integrating debiased machine learning and nonparametric modeling to enable robust inference under high-dimensional continuous or discrete covariates. Theoretically, we establish consistency, asymptotic normality, and double robustness of the estimator over general function classes. Our approach is the first to achieve falsifiable identification and efficient estimation of distributional causal effects under complex confounding structures, bridging theoretical rigor with empirical applicability.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Action, Change, and Causality

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

We introduce a novel extension of the influential changes-in-changes (CiC) framework [Athey and Imbens, 2006] to estimate the average treatment effect on the treated (ATT) and distributional causal estimands in panel data settings with unmeasured confounding. While CiC relaxes the parallel trends assumption inherent in difference-in-differences (DiD), existing approaches typically accommodate only a single scalar unobserved confounder and rely on monotonicity assumptions between the confounder and the outcome. Moreover, current formulations lack inference procedures and theoretical guarantees that accommodate continuous covariates. Motivated by the intricate nature of confounding in empirical applications and the need to incorporate continuous covariates in a principled manner, we make two key contributions in this technical report. First, we establish nonparametric identification under a novel set of assumptions that permit high-dimensional unmeasured confounders and non-monotonic relationships between confounders and outcomes. Second, we construct efficient estimators that are Neyman orthogonal to infinite-dimensional nuisance parameters, facilitating valid inference even in the presence of high-dimensional continuous or discrete covariates and flexible machine learning-based nuisance estimation.

Problem

• Extend CiC framework for ATT estimation with unmeasured confounding
• Address limitations of single scalar confounder and monotonicity assumptions
• Enable valid inference with high-dimensional continuous covariates

Innovation

• Extends CiC framework for ATT estimation
• Allows high-dimensional unmeasured confounders
• Uses Neyman orthogonal efficient estimators