📈 economics

Inference in Regression Discontinuity Designs with Clustered Data

This paper addresses the lack of theoretical attention to clustered sampling in regression discontinuity designs by introducing a general framework that establishes the asymptotic normality of standard estimators, critiques the limitations of current clustered standard errors, and proposes a novel nearest-neighbor variance estimator to improve inference accuracy.

Original authors: Claudia Noack, Tomasz Olma, Christoph Rothe

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Claudia Noack, Tomasz Olma, Christoph Rothe

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

Imagine you are a detective trying to solve a mystery: Does a specific event cause a change in behavior?

In the world of economics and social science, researchers often use a tool called Regression Discontinuity (RD). Think of it like a "cutoff line" on a map.

The Scenario: Imagine a scholarship is given to students who score 50% or higher on a test.
The Logic: A student who scores 49.9% is almost identical to a student who scores 50.1%. The only difference is the scholarship. By comparing these two groups, we can see if the scholarship caused better outcomes (like getting a job later).

This works beautifully when everyone is an individual, like a bag of mixed marbles where every marble is unique and independent.

The Problem: The "Group" Effect

However, in the real world, data isn't just a bag of marbles; it's often a bag of clusters.

Example: Instead of individual students, you might be looking at schools. All students in "School A" share the same teachers, the same lunch menu, and the same principal. They aren't independent; they are a team.
The Issue: If you treat students in the same school as if they were independent strangers, your detective work gets messy. You might think you have 1,000 pieces of evidence, but really, you only have 20 schools. This makes your conclusions shaky.

For a long time, statisticians didn't have a perfect rulebook for how to handle these "school" (cluster) situations in RD designs. They were using old tools that sometimes gave answers that were too confident (lying about how sure they were) or too cautious (missing real effects).

The Solution: A New Detective's Toolkit

The paper by Noack, Olma, and Rothe is like a new, specialized manual for detectives working with these "school" groups. They introduce two main innovations:

1. The "Traffic Light" System (Asymptotic Frameworks)

The authors realized that not all "schools" are the same size or structure.

Small Groups: Some clusters are tiny (like a few students).
Huge Groups: Some clusters are massive (like a whole city district).
The Analogy: Imagine trying to count cars at a traffic light.
- If the cars are scattered (independent), it's easy.
- If the cars are in convoys (clusters), the traffic light turns green for the whole convoy at once.
- The authors created four different "traffic rules" (frameworks) depending on how big the convoys are and how they move. They proved that no matter which rule applies, you can still get a valid answer if you follow the right math.

2. The "Neighborly" Standard Error (The CNN Method)

This is the paper's biggest invention. When you want to know how "sure" you are about your answer, you calculate something called a Standard Error. It's like a "margin of error" on a poll.

The Old Way (The Naive Approach): Imagine you want to guess the temperature. You look at your immediate neighbor. But if your neighbor is your brother, and you both live in the same house, his temperature is exactly the same as yours. If you use him to guess, you aren't getting new information; you're just repeating the same data. The old methods often did this by accident, leading to bad guesses.
The New Way (Clustered Nearest-Neighbors - CNN): The authors invented a smarter way to pick your "neighbors."
- The Rule: "To guess the temperature for House A, look at House B and House C, but never look at House A's brother."
- They ensure that the people you compare are from different groups (different schools, different cities). This guarantees that you are getting fresh, independent information.
- The Result: This new method (CNN) gives a much more accurate "margin of error." It stops you from being overconfident (thinking you know more than you do) or underconfident.

Why Does This Matter?

Think of it like baking a cake.

The RD Design is the recipe.
The Clusters are the ingredients (eggs, flour, sugar) that come in batches.
The Old Methods assumed every egg was unique. If you used a batch of eggs from the same carton, the cake might taste weird, and you wouldn't know why.
This Paper gives you a new way to taste the batter. It tells you exactly how to adjust your recipe when you know the eggs came in batches. It ensures that when you say, "This cake is delicious," you aren't just guessing because you tasted the same egg twice.

The Bottom Line

The authors have built a universal guide for researchers who study "cutoff" effects in a world where people live in groups. They proved that:

You can still get reliable answers even with grouped data, provided you follow their specific rules.
They invented a new, smarter calculator (the CNN standard error) that fixes the mistakes made by previous tools, ensuring that policy decisions (like who gets a scholarship or a new law) are based on solid, trustworthy math.

In short: They taught us how to count the votes correctly when the voters are sitting in teams.

1. Problem Statement

Regression Discontinuity (RD) designs are a cornerstone of causal inference in economics and social sciences. However, standard theoretical results for RD estimators (specifically local linear estimators) typically assume independent and identically distributed (i.i.d.) observations. In practice, empirical RD studies frequently involve clustered data (e.g., students within schools, voters within districts, firms within sectors), where observations within a cluster are dependent.

Despite the prevalence of clustered sampling (used in ~80% of recent RD studies), the theoretical literature lacks a unified framework for:

Establishing the asymptotic normality of local linear RD estimators under general clustering patterns (including growing cluster sizes).
Providing valid variance estimators that are consistent and efficient in finite samples. Existing methods, such as standard clustered regression residuals or naive adaptations of nearest-neighbor estimators, are shown to be either inconsistent or overly conservative.

2. Methodology and Framework

A. Model and Estimator

The authors consider a sharp RD design where treatment is assigned if a running variable $X_{gi}$ exceeds a cutoff (normalized to 0). The data consists of $G$ clusters, with $n_g$ units in cluster $g$ . The outcome model is:
$Y_{gi} = \mu(X_{gi}) + \varepsilon_{gi}$
where $\varepsilon_{gi}$ can be arbitrarily dependent within a cluster but independent across clusters.

The estimator of interest is the local linear RD estimator ( $\hat{\tau}(h)$ ), defined as the difference in local linear regression estimates on either side of the cutoff. This estimator is a weighted sum of outcomes: $\hat{\tau}(h) = \sum_{g,i} w_{gi}(h) Y_{gi}$ .

B. High-Level Conditions for Asymptotic Normality

The paper derives conditions under which the estimator is asymptotically normal. The core contribution is Assumption 1, which restricts the behavior of the weights $w_{gi}(h)$ relative to the conditional variance $se^2(h)$ . Specifically, it requires that no single cluster dominates the variance (no "large cluster" effect that breaks the CLT) and that the sum of absolute weight cross-products is bounded.

These high-level conditions translate into restrictions on cluster sizes within the estimation window ($nh$). The authors formalize this through four Asymptotic Frameworks:

Framework I & II (Small Effective Cluster Sizes): Suitable when the number of units per cluster within the bandwidth is small. Framework I assumes a bounded joint density of the running variable within clusters; Framework II allows for degenerate distributions (e.g., identical running variables within a cluster) but imposes stricter rate conditions on cluster sizes.
Framework III & IV (Large Effective Cluster Sizes): Relax the rate conditions on cluster sizes but require specific assumptions on the covariance structure of the residuals or the standard error itself. These cover settings where clusters are large even in a shrinking neighborhood of the cutoff.

C. Variance Estimation

The paper addresses the estimation of the conditional variance $se^2(h)$ .

Clustered Regression Residual (CRR): The standard approach using squared residuals. The authors prove this is consistent under the same conditions required for asymptotic normality. However, like in i.i.d. settings, it can be overly conservative (biased upward) in finite samples, especially when the conditional mean function $\mu(x)$ has high curvature.
Naive Clustered Nearest-Neighbors (Naive CNN): A direct adaptation of the i.i.d. nearest-neighbor variance estimator. The authors show this is invalid for clustered data because selecting neighbors without regard to cluster structure introduces bias in covariance estimation.
Proposed Clustered Nearest-Neighbors (CNN): A novel variance estimator designed specifically for clustered RD.
- Mechanism: For each observation $i$ in cluster $g$ , the method selects two sets of "companion" clusters ( $R^1_g, R^2_g$ ) that are disjoint from $g$ and from each other. Neighbors are chosen from these companion clusters.
- Key Innovation: By ensuring neighbors come from different clusters, the method exploits cross-cluster independence to construct an unbiased estimator of the covariance terms without relying on parametric assumptions about the error structure.

3. Key Contributions

Unified Asymptotic Theory: The paper provides the first general framework establishing the asymptotic normality of local linear RD estimators under arbitrary clustering patterns, including cases where cluster sizes grow with the sample size. It clarifies how the convergence rate of the variance depends on the interplay between the bandwidth, cluster sizes, and within-cluster dependence.
Novel Variance Estimator (CNN): The authors propose and prove the consistency of a Clustered Nearest-Neighbors (CNN) standard error. Unlike the CRR, the CNN estimator is shown to have lower finite-sample bias, particularly when the curvature of the regression function is substantial.
Diagnostic Tools: The paper introduces a practical "rule of thumb" (based on weight ratios $w_{max}$ and $w_{sum}$ ) to help applied researchers determine whether their data fits the "small cluster" frameworks (I/II) or requires the "large cluster" frameworks (III/IV).
Empirical Validation: The authors re-analyze four distinct empirical RD applications (covering disaster relief, elections, political participation, and OSHA penalties) to demonstrate that the CNN estimator performs robustly across different clustering structures, often yielding standard errors closer to the true sampling variability than the CRR or naive methods.

4. Key Results

Asymptotic Normality: Under high-level weight conditions (Assumption 1) and moment conditions (Assumption 2), the local linear RD estimator is asymptotically normal:
$\frac{\hat{\tau}(h) - E[\hat{\tau}(h)|X_n]}{se(h)} \xrightarrow{d} N(0, 1)$
Convergence Rates: The convergence rate of the variance $se^2(h)$ is generally slower than the i.i.d. rate of $(nh)^{-1}$ . It depends on a term $\lambda_n$ related to the sum of squared cluster sizes. In Frameworks I/II, the rate is $(nh)^{-1}$ ; in Frameworks III/IV, it can be slower, dominated by the cluster size structure.
Consistency of CNN: The proposed CNN standard error is consistent ( $\hat{se}^2_{CNN} / se^2 \xrightarrow{p} 1$ ) under the same high-level assumptions, provided the nearest neighbors are sufficiently close (Assumption 6).
Finite Sample Performance: In simulations and empirical applications, the CNN estimator reduces the upward bias of the CRR estimator. The "Naive CNN" is shown to be inconsistent, often underestimating variance when running variables are highly correlated within clusters.

5. Significance

This paper fills a critical gap in the econometric literature on RD designs. While cluster-robust inference is standard in linear regression, its application to nonparametric local polynomial RD estimators was theoretically underdeveloped.

For Theorists: It extends the theory of nonparametric estimation to clustered data, showing that localization (focusing on the cutoff) creates unique challenges regarding cluster sizes that differ from full-sample estimators.
For Practitioners: It offers a rigorous justification for using clustered standard errors in RD and, more importantly, provides a superior alternative (CNN) that improves inference accuracy in finite samples. The diagnostic tools allow researchers to assess whether their specific data structure supports standard asymptotic approximations or requires more robust variance estimation techniques.

In summary, Noack, Olma, and Rothe provide a comprehensive solution to the problem of inference in clustered RD designs, ensuring that researchers can draw valid causal conclusions even when data exhibits complex dependence structures.