Regression discontinuity design: estimating the treatment effect with standard parametric rate

Debarghya Mukherjee, Moulinath Banerjee, Ya’acov Ritov

Abstract

Regression discontinuity design (RDD) models are widely used for the assessment of treatment effects in psychology, econometrics, and biomedicine, specifically in situations where treatment is assigned to an individual based on their characteristics (e.g., a scholarship allocated based on merit) instead of being allocated randomly. Popular methods that have been largely employed to date for estimation of such treatment effects suffer from slow rates of convergence—specifically, slower than the standard $\sqrt{n}$ parametric rate. In this paper, we present a new model and method that allows estimation of the treatment effect at the $\sqrt{n}$ rate in the presence of fairly general forms of confoundedness. Moreover, we show that our estimator is semi-parametrically efficient in certain situations. We analyze two real datasets via our method and compare our results with those obtained by using previous approaches.

Introduction

In fields like economics and medicine, researchers frequently seek to quantify the impact of a specific "treatment," such as a policy intervention or a medical protocol. While Randomized Clinical Trials (RCTs) are the gold standard, they are often impractical or unethical. In these cases, researchers turn to Regression Discontinuity Design (RDD), where treatment is assigned based on whether an individual’s "running variable" (like a test score or GPA) crosses a specific threshold.

Traditionally, RDD analysis relies on a "local" approach, focusing only on observations immediately surrounding the cutoff. The logic is that individuals just above and just below the threshold are nearly identical, creating a quasi-experimental setting. However, this approach discards the vast majority of the dataset, leading to a statistical "speed limit": the convergence rate of the estimator is significantly slower than the standard $\sqrt{n}$ rate. This paper introduces a "global" approach that leverages the entire dataset to achieve parametric efficiency.

Methodological Framework

The core innovation of this research is the augmentation of the standard RDD outcome model with a secondary model for the treatment assignment variable. By accounting for the relationship between background covariates and the running variable, the authors can mitigate the bias that usually necessitates a local-only focus.

The Model

The system is represented by two primary equations:

  1. The Outcome Equation ($Y$): This models the result of interest, incorporating a non-parametric function $b(\eta)$ to account for endogeneity—unobserved factors that influence both the assignment score and the outcome.
  2. The Assignment Equation ($Q$): This models how the assignment score (the running variable) relates to observed covariates.

Data Splitting and Estimation

To maintain mathematical rigor and prevent over-fitting, the authors employ a three-part data-splitting strategy:

  • Partition 1: Used to estimate the relationship between covariates and the assignment score.
  • Partition 2: Employs B-splines (flexible mathematical curves) to estimate the derivative of the unknown "nuisance" function capturing endogeneity.
  • Partition 3: Integrates these estimates into a final regression to isolate the treatment effect ($\alpha_0$).

To ensure no data is wasted, the roles of these partitions are rotated, and the resulting estimates are averaged. This transformation effectively converts a complex RDD problem into an approximate partial linear model, a well-understood framework in statistical theory.

Empirical Applications

The authors validated their method using two distinct real-world datasets, highlighting the practical implications of their high-precision estimator.

Turkey Municipality Voting Data

This study investigated whether the election of Islamic mayors in 1994 influenced women’s high school education rates by the year 2000 ($n = 2629$). Traditional local RDD methods previously suggested a significant positive effect. However, applying the authors' global method yielded a point estimate of 0.407 with a 95% Bootstrap Confidence Interval of (-0.423, 1.419). Because this interval includes zero, the authors concluded that there was no statistically significant effect, suggesting that traditional local methods may be overly sensitive to bandwidth selection or local noise.

Canadian University GPA Data

This study examined whether placing students on academic probation based on a GPA cutoff influenced subsequent academic performance. The global method produced a point estimate of 0.273 with a 95% Confidence Interval of (0.248, 0.315). This indicates a significant positive effect, supporting the hypothesis that probation serves as an effective incentive for students to improve their academic standing.

Key Insights and Theoretical Contributions

  • Breaking the Speed Limit: The authors provide mathematical proof that their estimator converges at the $\sqrt{n}$ rate and follows an asymptotic normal distribution. This allows for more precise inferences than previously possible in RDD settings.
  • Semi-parametric Efficiency: Under certain conditions, such as normally distributed errors, the estimator is proven to be semi-parametrically efficient, meaning it achieves the lowest possible variance among a broad class of estimators.
  • Global Utility: By modeling the running variable, researchers can move beyond the cutoff, utilizing the full power of their collected data without sacrificing the validity of the treatment effect estimate.

Limitations and Future Research

While this method represents a significant advancement, the authors note several areas for future exploration:

  • Error Assumptions: The current framework assumes errors are independent of covariates, which may not hold in highly complex environments.
  • High-Dimensionality: Future work could adapt this model for "Big Data" contexts where the number of covariates ($p$) is large relative to the sample size ($n$).
  • Non-linear Treatment Effects: The current model assumes a constant treatment effect (a uniform shift). Extending the method to account for treatment effects that vary non-linearly with the assignment score would further increase its flexibility.

Summary updated

Abstract

Regression discontinuity design models are widely used for the assessment of treatment effects in psychology, econometrics and biomedicine, specifically in situations where treatment is assigned to an individual based on their characteristics (e.g. scholarship is allocated based on merit) instead of being allocated randomly, as is the case, for example, in randomized clinical trials. Popular methods that have been largely employed till date for estimation of such treatment effects suffer from slow rates of convergence (i.e. slower than $\sqrt{n}$). In this paper, we present a new model and method that allows estimation of the treatment effect at $\sqrt{n}$ rate in the presence of fairly general forms of confoundedness. Moreover, we show that our estimator is also semi-parametrically efficient in certain situations. We analyze two real datasets via our method and compare our results with those obtained by using previous approaches. We conclude this paper with a discussion on some possible extensions of our method.
BibTeX
@misc{mukherjee2021regression,
  author = {Debarghya Mukherjee and Moulinath Banerjee and Ya’acov Ritov},
  title = {{Regression discontinuity design: estimating the treatment effect with standard parametric rate}},
  year = {2021},
  eprint = {2102.11229},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.11229},
}
Figure