Persuasion Effects in Regression Discontinuity Designs

Sung Jae Jun, Sokbae LeearXiv 2025 · DOI · arXiv

Abstract

We develop a framework for identifying and estimating persuasion effects in regression discontinuity (RD) designs. The RD persuasion rate measures the probability that individuals at the threshold would take the action if exposed to a persuasive message, given that they would not take the action without exposure. We present identification results for both sharp and fuzzy RD designs, derive sharp bounds under various data scenarios, and extend the analysis to local compliers. Estimation and inference rely on local polynomial regression, enabling straightforward implementation with standard RD tools. Applications to public health and media illustrate its empirical relevance.

Introduction

In social science and policy analysis, a critical question is how effectively a specific message—such as a public health campaign, a political advertisement, or a news broadcast—actually changes behavior. This is typically quantified by the persuasion rate: the percentage of the "unconverted" population that takes an action only after being exposed to the message.

While methods for calculating persuasion rates are well-established for randomized controlled trials (RCTs), they have remained underdeveloped for Regression Discontinuity (RD) designs. RD designs are a staple of quasi-experimental research, where treatment is assigned based on a "running variable" crossing a specific threshold (e.g., an age cutoff for voting or a geographic boundary for a policy). This paper fills this methodological gap, providing a formal framework to identify and estimate persuasion effects at the RD threshold.

The RD Persuasion Framework

The authors introduce the RD persuasion rate ($\theta$), defined specifically at the cutoff point. This metric is designed to filter out "preaching to the converted"—it ignores individuals who would have taken the action regardless of the message and focuses purely on the "swing" population at the margin.

Identification in Sharp and Fuzzy Designs

The framework addresses two primary RD scenarios:

  • Sharp RD Designs: In these settings, treatment is a deterministic function of the running variable (e.g., everyone above the threshold is exposed). Here, the persuasion rate is point identified, meaning it can be calculated directly from the observed jump in outcomes at the threshold.
  • Fuzzy RD Designs: In many real-world cases, the threshold only changes the probability of treatment. The paper derives "sharp bounds"—the tightest possible mathematical range—for the persuasion rate. These bounds are adaptable to different data environments, including scenarios where researchers only have access to aggregate-level data rather than individual-level joint data (knowing exactly who saw the message and how they subsequently acted).

Estimation and Inference

To make the framework accessible to practitioners, the authors utilize local polynomial regression. This allows the RD persuasion rate to be estimated using standard RD software. Furthermore, the paper provides methods for bias-corrected inference and confidence intervals, which are essential when dealing with the partial identification (ranges) inherent in fuzzy designs.

Empirical Applications

The framework's utility is demonstrated through two distinct applications in public health and political science.

Using data surrounding the 1905 U.S. Supreme Court ruling in Jacobson v. Massachusetts, which upheld vaccine mandates, the authors examined the persuasive power of the resulting media discourse. In this sharp RD application, the conventional RD estimate suggested a 6.3% persuasion rate toward anti-vaccine sentiment. However, applying the paper’s robust bias-corrected estimation revealed a significantly higher persuasion rate of 14.3%, suggesting a substantial "backlash" effect triggered by the legal mandate.

2. Media Influence in Italian Politics

The authors also analyzed the rollout of digital TV in Italy, which reduced exposure to media slanted toward Silvio Berlusconi. This fuzzy RD application used aggregate town-level data. The framework established a lower bound for the persuasion rate of approximately 10.4%, representing the rate at which voters were persuaded not to vote for Berlusconi’s coalition due to the change in media access.

Key Insights

  • Monotonicity as a Safeguard: The model typically assumes a "Monotone Treatment Response" (the message does not cause the opposite of the intended effect). However, the authors prove that even if this assumption is violated, the RD estimate still functions as a robust lower bound on the true effect.
  • Data Flexibility: A major practical contribution is the proof that individual-level data is not strictly necessary to derive useful bounds. This is particularly valuable in political science, where voting behavior is often only observable at the aggregate level.
  • Importance of Bias Correction: The empirical results highlight that without proper bias correction, researchers may significantly underestimate the actual power of persuasive communication, especially in smaller samples.

Limitations and Future Directions

While the current framework provides a robust toolset for binary outcomes (e.g., voting vs. not voting), the authors identify several areas for expansion:

  • Continuous Outcomes: Extending the logic to non-binary actions.
  • Complex Treatments: Adapting the model for multiple running variables or varying levels of message intensity.
  • Heterogeneity: Integrating machine learning to identify how persuasion effects differ across various demographic or socioeconomic subpopulations.

By providing a "plug-and-play" methodology for RD designs, this research enables a more nuanced understanding of how information shapes behavior at the margins of policy and law.

Summary updated

Abstract

We develop a framework for identifying and estimating persuasion effects in regression discontinuity (RD) designs. The RD persuasion rate measures the probability that individuals at the threshold would take the action if exposed to a persuasive message, given that they would not take the action without exposure. We present identification results for both sharp and fuzzy RD designs, derive sharp bounds under various data scenarios, and extend the analysis to local compliers. Estimation and inference rely on local polynomial regression, enabling straightforward implementation with standard RD tools. Applications to public health and media illustrate its empirical relevance.
BibTeX
@misc{jun2025persuasion,
  author = {Sung Jae Jun and Sokbae Lee},
  title = {{Persuasion Effects in Regression Discontinuity Designs}},
  year = {2025},
  eprint = {2509.26517},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.26517},
  howpublished = {arXiv},
}
Figure