Identification and Debiased Learning of Causal Effects with General Instrumental Variables

Shuyuan Chen, Peng Zhang, Yifan CuiarXiv 2025 · DOI · arXiv

Abstract

Instrumental variable methods are fundamental to causal inference when treatment assignment is confounded by unobserved variables. In this article, we develop a general nonparametric causal framework for identification and learning with multi-categorical or continuous instrumental variables. Specifically, the mean potential outcomes and the average treatment effect can be identified via a regular weighting function derived from the proposed framework. Leveraging semiparametric theory, we derive efficient influence functions and construct two consistent, asymptotically normal estimators via debiased machine learning. The first estimator uses a prespecified weighting function, while the second estimator selects the optimal weighting function adaptively. Extensions to longitudinal data, dynamic treatment regimes, and multiplicative instrumental variables are further developed. We demonstrate the proposed method by employing simulation studies and analyzing real data from the Job Training Partnership Act program.

Introduction

In causal inference, the primary challenge is often unobserved confounding—hidden factors that influence both the treatment (e.g., participation in a program) and the outcome (e.g., future earnings). While Instrumental Variable (IV) methods have long served as a solution by utilizing a third variable that affects the treatment but has no direct link to the outcome, traditional approaches are often limited. Most rely on rigid linear models, such as Two-Stage Least Squares, or are restricted to binary instruments. Furthermore, many nonparametric methods require a "monotonicity" assumption—the requirement that the instrument affects all individuals in the same direction—which is frequently violated in practice.

This paper introduces a flexible, nonparametric framework for "General Instrumental Variables" (GIV). By moving beyond binary constraints and restrictive assumptions, the authors provide a robust methodology for estimating the Average Treatment Effect (ATE) across diverse data types, including multi-categorical and continuous instruments.

The General Instrumental Variable Framework

The proposed framework rests on two primary theoretical pillars designed to ensure the identification of causal effects in complex settings:

  1. Regular Weighting Functions (RWF): The authors introduce the RWF as a mechanism to identify mean potential outcomes. By weighting observed data, the RWF utilizes the covariance between the instrument and the treatment to effectively "filter out" bias introduced by hidden confounders.
  2. The Additive IV (AIV) Condition: To facilitate nonparametric identification, the framework employs the AIV condition. This assumes that the instrument's effect on the treatment is additive relative to hidden confounders. This serves as a natural generalization of "no-interaction" assumptions found in simpler models, providing a sufficient condition for identifying causal effects without requiring monotonicity.

Estimation via Debiased Machine Learning

For practical estimation, the paper leverages Debiased Machine Learning (DML). The authors derive Efficient Influence Functions (EIFs)—mathematical tools that dictate how to update an estimate to achieve maximum accuracy and robustness. Based on these functions, two distinct estimators are proposed:

  • Prespecified Estimator: Utilizes a fixed, pre-determined weighting function.
  • Adaptive Estimator: Automatically selects the "optimal" weighting function from the data. By treating the optimal instrument as a nuisance function and estimating it via machine learning, this estimator minimizes variance and achieves the semiparametric efficiency bound.

Beyond static treatments, the framework extends to longitudinal data and dynamic treatment regimes. The authors develop a "backward cross-fitting" procedure to handle sequences of treatments and instruments over time, making the method applicable to complex, multi-stage clinical or economic studies.

Empirical Validation and Results

The framework was validated through a combination of synthetic simulations and real-world data analysis:

  • Simulation Studies: Testing various Data Generating Processes (DGPs) confirmed that the DML approach maintains low bias and achieves near-perfect 95% coverage rates. The adaptive estimator consistently outperformed prespecified weights in terms of precision.
  • Job Training Partnership Act (JTPA): Analyzing this program revealed that job training is significantly more effective for individuals in lower-earning brackets than for those in higher ones, highlighting the model's utility for targeted policy design.
  • CollegeDistance Dataset: Using distance to college as an instrument for educational attainment, the framework demonstrated its ability to handle continuous instruments to estimate the impact of education on subsequent income.

Conclusion

This research provides a unified perspective on instrumental variable relevance through the lens of Regular Weighting Functions. By bypassing the need for monotonicity and allowing for continuous or multi-categorical instruments, the framework offers a more versatile toolset for researchers. While the method relies on the AIV or Multiplicative IV (MIV) conditions—which must be justified by domain knowledge—it significantly weakens the parametric constraints of traditional IV analysis. Future work may further explore the stability of these estimators in the presence of "weak instruments" or in extremely high-dimensional settings where the number of confounders is vast.

Summary updated

Abstract

Instrumental variable methods are fundamental to causal inference when treatment assignment is confounded by unobserved variables. In this article, we develop a general nonparametric causal framework for identification and learning with multi-categorical or continuous instrumental variables. Specifically, the mean potential outcomes and the average treatment effect can be identified via a regular weighting function derived from the proposed framework. Leveraging semiparametric theory, we derive efficient influence functions and construct two consistent, asymptotically normal estimators via debiased machine learning. The first estimator uses a prespecified weighting function, while the second estimator selects the optimal weighting function adaptively. Extensions to longitudinal data, dynamic treatment regimes, and multiplicative instrumental variables are further developed. We demonstrate the proposed method by employing simulation studies and analyzing real data from the Job Training Partnership Act program.
BibTeX
@misc{chen2025identification,
  author = {Shuyuan Chen and Peng Zhang and Yifan Cui},
  title = {{Identification and Debiased Learning of Causal Effects with General Instrumental Variables}},
  year = {2025},
  eprint = {2510.20404},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.20404},
  howpublished = {arXiv},
}
Figure