📈 economics

PLRD: Partially Linear Regression Discontinuity Inference

This paper introduces the Partially Linear Regression Discontinuity (PLRD) estimator, a new method that demonstrates substantially lower estimation error and valid, shorter confidence intervals compared to existing approaches in regression discontinuity designs, validated through a novel simulation framework using Wasserstein generative adversarial networks to replicate real-world data.

Original authors: Aditya Ghosh, Guido Imbens, Stefan Wager

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditya Ghosh, Guido Imbens, Stefan Wager

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of economic research, scientists often face a difficult puzzle: how to measure the true effect of a policy or program when they cannot run a controlled experiment. Imagine a government decides to give extra funding to schools only if their test scores fall below a certain line. Researchers want to know if that money actually improved student outcomes. They cannot simply compare schools that got the money to those that didn't, because the schools that received funding were likely struggling for other reasons before the money arrived. To solve this, economists use a tool called a regression discontinuity design. This method treats the cutoff line as a natural experiment. It assumes that schools just barely below the line are nearly identical to schools just barely above it, except for the fact that one group received the treatment and the other did not. By looking very closely at the data right around that dividing line, researchers can isolate the impact of the policy.

For decades, the standard way to draw conclusions from this data has relied on fitting a simple curve to the points near the line. However, this approach has a hidden flaw. The curve often bends in ways the simple model cannot see, leading researchers to draw conclusions that are too confident or simply wrong. The confidence intervals—the range of values where the true answer is likely to hide—often fail to capture the real effect, even when the researchers think they are being careful. This uncertainty makes it hard to know if a policy is truly working or if the results are just a statistical illusion.

A team of researchers at Stanford University has proposed a new way to handle this problem, offering a more reliable method for these critical calculations. They call their approach partially linear regression discontinuity inference. Instead of assuming the data follows a simple, rigid shape, their method allows the relationship between the running variable and the outcome to be more flexible, while still imposing a reasonable structure on how the treatment effect changes. To test if this new method actually works better than the old ones, the researchers did not just rely on theoretical math. They built a sophisticated simulation engine using artificial intelligence. They fed the engine real data from twelve famous studies that had previously used this type of design, ranging from the effects of incumbency in US elections to the impact of summer school on student grades. The AI learned the patterns and quirks of these real-world datasets and then generated thousands of synthetic versions that looked and behaved exactly like the original data, but where the researchers knew the true answer.

When they ran the new method against the standard approaches on these synthetic datasets, the results were striking. The traditional methods, which are widely used today, frequently failed to include the true answer in their calculated ranges, often missing it by a significant margin. The new method, however, consistently found the correct answer within its confidence intervals. More importantly, it did not have to sacrifice precision to achieve this accuracy. The ranges produced by the new method were not only valid but also much narrower than those produced by other reliable, conservative methods. In many cases, the new approach reduced the error in the estimates by a large margin, providing a clearer picture of the policy's true impact.

The researchers also checked how their method held up when the data was messy or when the underlying rules of the system changed, such as when the running variable was not a smooth number but a discrete score. Even in these difficult scenarios, the new method adapted and continued to perform well, whereas other methods struggled or became overly cautious, producing ranges so wide they were useless for decision-making. The key to their success was a two-step process. First, they used a preliminary analysis to understand how much the data curved, essentially measuring the "roughness" of the relationship. Then, they used this measurement to build a safety net that accounted for the worst-case scenario of that curvature, without assuming the worst case was actually happening. This allowed them to be precise when the data was smooth and safe when it was not.

This work suggests that the tools economists have been using for years may be leaving valuable information on the table or, worse, leading them astray. By combining a flexible model with a rigorous way of accounting for uncertainty, the new method offers a path to more trustworthy conclusions. It does not require researchers to guess the right settings or make arbitrary choices; the method automatically adjusts to the data it sees. While the findings are based on simulations that mimic real-world conditions rather than a single new real-world application, the consistency of the results across twelve different high-profile studies gives strong reason to believe the approach is robust. For policymakers and researchers trying to understand the real effects of interventions, this new tool promises to turn the blurry picture of regression discontinuity designs into something much sharper and more reliable.

Technical Summary: Partially Linear Regression Discontinuity Inference (PLRD)

Problem Statement

Regression Discontinuity Designs (RDDs) are a dominant research design in empirical economics, yet standard approaches to constructing confidence intervals often exhibit suboptimal behavior in practice. Specifically, widely used local-polynomial methods frequently fail to achieve nominal coverage in realistic settings. Existing "bias-corrected" methods (e.g., rdrobust) often suffer from under-coverage, while "bias-aware" methods (e.g., RDHonest) achieve reliable coverage but at the cost of excessively wide confidence intervals. The central question addressed by this paper is whether it is possible to design automated confidence intervals that reliably achieve nominal coverage in realistic settings while remaining competitive with conventional intervals in terms of width.

Methodology: The PLRD Estimator

The authors propose the Partially Linear Regression Discontinuity (PLRD) estimator. This method integrates the "bias-aware" paradigm (minimax linear estimation) with specific structural assumptions about the treatment effect heterogeneity.

Core Assumptions

The PLRD method relies on two primary assumptions:

Smoothness (Assumption 1): The conditional response functions $\mu_w(x)$ are three times continuously differentiable. This is a standard smoothness assumption consistent with bias-corrected literature.
Partial Linearity (Assumption 2): The conditional average treatment effect (CATE) function, $\tau(x) = \mu_1(x) - \mu_0(x)$ , has a linear dependence on the running variable $x$ . Specifically, $\tau(x) = \tau_c + \beta(x - c)$ for some $\beta \in \mathbb{R}$ . This assumption posits that the treatment effect varies more simply than the baseline outcome.

Algorithmic Procedure

The PLRD procedure (Algorithm 1) proceeds as follows:

Curvature Estimation: Under Assumption 1, the method estimates a Lipschitz constant $B$ for the second derivatives of the response functions. This is achieved via a local cubic regression on a pilot bandwidth.
Minimax Optimization: The method constructs a set of all possible response functions consistent with the estimated $B$ -Lipschitz curvature bound and the linear treatment effect assumption (Assumption 2). Within this class, it uses numerical convex optimization (quadratic programming) to derive a minimax linear estimator for the treatment effect at the cutoff, $\tau_c$ .
Bias-Aware Inference: The method identifies the worst-case bias of the estimator within the specified class and constructs confidence intervals that simultaneously account for the standard error and this potential bias.
Robustness Safeguard: To protect against violations of Assumption 2, the procedure includes an initial ANOVA test comparing restricted and unrestricted cubic models. If the linear treatment effect assumption is rejected, the method falls back to a more flexible specification that only relies on Assumption 1.

Theoretical Guarantees

The paper establishes that PLRD yields asymptotically valid confidence intervals under the stated assumptions. Specifically, Theorem 1 proves that for any significance level $\alpha$ , the probability that the true parameter $\tau_c$ lies within the PLRD interval converges to at least $1-\alpha$ as the sample size $n \to \infty$ . The proof relies on a central limit theorem for the PLRD estimator and demonstrates that the learned bias bound is conservative relative to the true bias.

Empirical Evaluation

The authors evaluate PLRD against three benchmarks: conventional local linear regression (with Imbens–Kalyanamaran bandwidth), the bias-corrected rdrobust method, and the bias-aware RDHonest method.

Simulation Design

To address the lack of ground truth in real-world data, the authors employ a semi-synthetic evaluation framework using Wasserstein Generative Adversarial Networks (WGANs). They train WGANs on twelve high-profile RDD applications (including Lee, 2008; Cattaneo et al., 2015; Ludwig and Miller, 2007; etc.) to generate synthetic datasets that mimic the moments and idiosyncrasies of the original data without enforcing parametric smoothness assumptions.

Key Results

Across 10,000 simulation replications for each of the twelve applications:

Coverage: PLRD achieves nominal (near-exact) 95% coverage across all settings. In contrast, conventional methods systematically undercover, and rdrobust achieves better but imperfect coverage (nominal rates in only 5 of 10 settings). RDHonest achieves reliable coverage but is often overly conservative.
Interval Width: PLRD produces confidence intervals that are substantially narrower than RDHonest and competitive with (often narrower than) rdrobust.
Estimation Error: PLRD demonstrates the lowest Root Mean Squared Error (RMSE) across all settings, outperforming all comparison methods.
Discrete Variables: The method performs robustly even when the running variable is discrete (e.g., test scores), a scenario where formal guarantees for the pilot-based curvature estimation are weaker.

Significance and Claims

The paper claims that PLRD provides an affirmative answer to the pragmatic question of whether automated, narrow, and valid confidence intervals are possible in realistic RDD settings.

Methodological Contribution: PLRD bridges the gap between the "bias-corrected" and "bias-aware" literatures. It adopts the minimax linear estimation framework of the bias-aware paradigm but utilizes the data-driven, asymptotic approach of the bias-corrected paradigm.
Role of Assumptions: The authors emphasize that the efficiency gains of PLRD are a joint consequence of both the higher-order smoothness (Assumption 1) and the partial linearity of the treatment effect (Assumption 2). Ablation studies confirm that relaxing either assumption leads to more conservative intervals.
Evaluation Framework: The paper advocates for the use of WGAN-based semi-synthetic simulations as a credible template for evaluating new econometric methods, arguing that they provide a neutral ground for comparing methods motivated by different assumptions.

The authors note that their approach relies on the assumption that the real datasets used for calibration are not "adversarial" in the sense that they could be indistinguishable from data where standard estimators fail. They encourage further exploration of sensitivity to different generative models.