Identifying average causal effect in regression discontinuity design with auxiliary data

Xiao Zhou Feng, Wenjie Hu, Pu Yang, Tingyu Li, Xiao‐Hua ZhouarXiv 2024 · DOI · arXiv

TL;DR

Regression Discontinuity Designs (RDD) are great for finding causal effects at a specific threshold, but they usually can't tell you how a treatment works for the general population. This paper introduces a novel framework that uses an auxiliary dataset to "bridge" the gap, allowing researchers to estimate the Average Treatment Effect (ATE) across the entire population while maintaining the rigor of a doubly robust estimator.

What Problem Does This Paper Address?

In the world of causal inference, the Regression Discontinuity Design (RDD) is a gold standard for "natural experiments." Imagine a scholarship given only to students who score above 90 on an exam. By comparing students who scored 89 to those who scored 91, we can estimate the effect of the scholarship because those students are essentially identical except for the treatment.

The catch? RDD only identifies the Local Average Treatment Effect (LATE)—the effect right at the 90-point cutoff. It tells us nothing about the student who scored 60 or 100. The fundamental problem is a violation of "positivity": at any given score, the treatment is deterministic. You either got the scholarship or you didn't. To find the Global Average Treatment Effect (ATE), you usually have to make incredibly strong (and often unrealistic) functional form assumptions.

The authors tackle this "extrapolation" challenge by asking: What if the score (the running variable) is just a proxy for a deeper, latent variable, and we have a second dataset that connects the two?

Methods

The authors propose an Auxiliary Variable Framework. They assume there exists an auxiliary variable ($U$) such that, once you control for $U$, the running variable ($X$) and the treatment ($W$) no longer provide additional information about the potential outcomes.

To make this work, they require two datasets:

  1. The Main Dataset: The standard RDD data $(X, W, Y)$, where $U$ is unobserved.
  2. The Auxiliary Dataset: A separate sample containing $(U, X)$, allowing the model to learn the relationship between the latent variable and the running variable.

The technical "secret sauce" involves Bridge Functions. Borrowing from proximal causal inference literature, the authors define an "outcome bridge function" (similar to an outcome regression) and a "treatment bridge function" (similar to an inverse propensity score).

They introduce three estimators:

  • An outcome-based estimator.
  • An inverse propensity weighted (IPW) estimator.
  • A Doubly Robust (DR) estimator, which remains consistent as long as either the outcome bridge or the treatment bridge is correctly specified.

Datasets

The paper evaluates the framework using two primary sources:

  • Simulated Data: Two settings were tested. Setting 1 used a binary auxiliary variable and a continuous outcome; Setting 2 used a continuous auxiliary variable and a binary outcome. These simulations verified that the estimators converge as sample sizes increase.
  • The Vitamin A ASD Study: A real-world application involving a program at the Children’s Hospital of Chongqing Medical University.
  • Main Dataset: 149 children in an RDD setup where Vitamin A was assigned based on serum retinol levels.
  • Auxiliary Dataset: 378 children where both serum retinol (the running variable) and retinoic acid (the auxiliary variable) were measured simultaneously.

Key Insights and Ideas

  • The Surrogate Solution: The core "aha" moment is treating the running variable as a surrogate. In the medical example, serum retinol is the "running variable" used for assignment, but retinoic acid is the "auxiliary variable" that actually participates in physiological activities. By linking them, we can bypass the deterministic nature of the RDD.
  • Data Fusion as Extrapolation: Instead of guessing how a curve looks far away from a cutoff, the authors use a second data source to provide the missing "overlap" needed for causal identification.
  • Neural Network Implementation: To solve the complex minimax optimization problems required for the bridge functions, the authors successfully employed two-layer neural networks, showing the framework is compatible with modern machine learning tools.

Key Results

  • Consistency and Robustness: In simulations, the Doubly Robust estimator lived up to its name. Even when one of the bridge functions was intentionally "misspecified" (using a bad model), the DR estimator still produced accurate results, whereas the single-model estimators failed.
  • Vitamin A and Autism: Applying the method to the ASD study, the authors found that Vitamin A supplementation had a positive effect on reducing the severity of symptoms (measured by Social Reaction Scale scores).
  • Statistical Significance: While the effect was numerically positive, it did not reach statistical significance. This highlights a practical reality: while the method provides a way to estimate the ATE, the resulting confidence intervals may be wide if the underlying sample size is small.

Limitations and Future Work

The authors acknowledge a few hurdles:

  • The "Good" Auxiliary Variable: The entire framework relies on finding a variable $U$ that satisfies the "latent confounding" and "latent positivity" assumptions. If $U$ is just as deterministic as $X$, the method fails.
  • Bridge Function Existence: While they provide theoretical conditions for the existence of these functions, in practice, ensuring they are well-posed can be difficult.
  • Generalization: The authors suggest that this "data fusion" approach could be extended beyond RDD to other observational data settings where the positivity assumption (the requirement that everyone has a non-zero chance of being in either group) is violated.

Summary updated

Abstract

Regression discontinuity designs are widely used when treatment assignment is determined by whether a running variable exceeds a predefined threshold. However, most research focuses on estimating local causal effects at the threshold, leaving the challenge of identifying treatment effects away from the cutoff largely unaddressed. The primary difficulty in this context is that the treatment assignment is deterministically defined by the running variable, violating the commonly assumed positivity assumption. In this paper, we introduce a novel framework for identifying the average causal effect in regression discontinuity designs. Our approach assumes the existence of an auxiliary variable for which the running variable can be seen as a surrogate, and an additional dataset that consists of the running variable and the auxiliary variable alongside the traditional regression discontinuity design setup. Under this framework, we propose three estimation methods for the ATE, which resembles the outcome regression, inverse propensity weighted and doubly robust estimators in classical causal inference literature. Asymptotically valid inference procedures are also provided. To demonstrate the practical application of our method, simulations are conducted to show the good performance of our methods; besides, we use the proposed methods to assess the causal effects of vitamin A supplementation on the severity of autism spectrum disorders in children, where a positive effect is found but with no statistical significance.
BibTeX
@misc{feng2024identifying,
  author = {Xiao Zhou Feng and Wenjie Hu and Pu Yang and Tingyu Li and Xiao‐Hua Zhou},
  title = {{Identifying average causal effect in regression discontinuity design with auxiliary data}},
  year = {2024},
  eprint = {2412.20840},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.20840},
}
Figure