Boundary Discontinuity Designs: Theory and Practice

1. Problem Statement

The Regression Discontinuity (RD) design is a leading causal inference method where treatment is assigned based on a univariate score crossing a scalar cutoff. However, many real-world policy assignments rely on multivariate scores (e.g., geographic coordinates, multiple exam scores) and boundary curves rather than a single point. This setup is known as the Boundary Discontinuity (BD) design (also called Multi-Score RD or Geographic RD).

While widely used, the BD design presents unique theoretical and practical challenges:

• Multidimensionality: The assignment boundary is a curve (1D submanifold) in a 2D space, complicating the definition of "closeness" and localization compared to univariate RD.
• Empirical Heterogeneity: Most empirical studies collapse the 2D problem into a 1D problem by using the shortest distance to the boundary, potentially masking heterogeneous treatment effects along the boundary.
• Methodological Gaps: There has been a lack of formal theoretical justification for the most common empirical strategies (pooling methods), particularly regarding identification, bias properties, and valid inference under complex boundary geometries.

2. Methodology and Framework

The paper synthesizes over 80 empirical papers and provides a rigorous theoretical framework for BD designs. The methodology is divided into two main approaches:

A. Pooling-Based Methods (The Dominant Empirical Strategy)

Most applied researchers (approx. 93% of the reviewed literature) use a pooling approach.

• Mechanism: They calculate the signed shortest distance ($D_i$) from each unit's location ($X_i$) to the boundary curve ($B$). They then restrict the sample to a "tubular neighborhood" where $|D_i| \leq h$ (bandwidth).
• Estimation: They treat $D_i$ as a univariate running variable and estimate a single Boundary Average Treatment Effect (BATE) using local polynomial regression.
• Common Specifications: The paper categorizes empirical specifications into eight types, ranging from simple difference-in-means to local linear regressions with boundary-segment fixed effects.
  • Critique: Many studies omit the crucial interaction term between the treatment indicator and the polynomial expansion of the score, leading to potential misspecification bias.
  • Inference Issue: Standard confidence intervals often fail because they ignore the bias introduced by using Mean Squared Error (MSE)-optimal bandwidths.

B. Heterogeneity and Aggregation Methods (The Proposed General Approach)

The authors advocate for analyzing the Boundary Average Treatment Effect Curve (BATEC), denoted as $\tau(x)$, which represents the treatment effect at every specific point $x$ along the boundary.

• Distance-Based Methods: Estimate $\tau(x)$ at specific grid points on the boundary by treating the distance to that specific point as the running variable.
  • Limitation: Requires the boundary to be smooth to achieve higher-order bias reduction; otherwise, bias remains high.
• Location-Based Methods (The Robust Solution): Use local bivariate regression directly on the coordinates $X_i$ relative to a grid point $b_j$ on the boundary.
  • Specification: Regress $Y_i$ on $T_i$, a polynomial expansion of $(X_i - b_j)$, and their interaction, weighted by a bivariate kernel.
  • Advantage: This approach does not require the boundary to be smooth for higher-order debiasing and captures the full geometric complexity of the assignment rule.

3. Key Contributions

1. Formal Identification and Asymptotics for Pooling:

   • The paper (citing Cattaneo et al., 2026) provides the first formal identification results for pooling-based BD methods using geometric measure theory.
   • It establishes that the probability limit of pooling estimators converges to a density-weighted average of treatment effects along the boundary ($\tau = \frac{\int_B \tau(x)f(x)dx}{\int_B f(x)dx}$).
   • It derives the bias and variance properties, showing that the convergence rate is similar to 1D nonparametric estimation due to the aggregation along the boundary.

2. Rigorous Inference Procedures:

   • The authors demonstrate that standard confidence intervals are invalid when using MSE-optimal bandwidths due to non-negligible bias.
   • They recommend Robust Bias Correction (RBC): A two-step procedure where the point estimate uses a lower-order polynomial (e.g., $p=1$) with MSE-optimal bandwidth, while the variance and confidence intervals are constructed using a higher-order polynomial (e.g., $q=2$) to correct for bias.

3. Framework for Heterogeneity:

   • The paper formalizes the BATEC ($\tau(x)$) as the fundamental building block for BD designs.
   • It defines aggregate parameters derived from the BATEC, such as the Weighted BATE (WBATE) and the Largest BATE (LBATE), allowing researchers to evaluate targeted policies or find optimal locations for intervention.

4. Empirical Taxonomy and Review:

   • The paper reviews 82 empirical papers across Economics, Political Science, and Education, documenting that the vast majority rely on pooling methods and often neglect heterogeneity or use suboptimal specifications (e.g., missing interaction terms).

4. Key Results

• Bias-Variance Trade-off: In pooling methods, the bias depends on the geometry of the boundary. If the boundary is irregular (common in geographic applications), standard univariate RD assumptions may not hold without specific geometric regularity conditions (e.g., rectifiable curves).
• Specification Sensitivity: Empirical specifications that fail to include the interaction between the treatment indicator and the score polynomial (e.g., $T_i \cdot r_p(D_i)$) suffer from larger misspecification bias compared to those that do.
• Validity of Inference: Standard inference fails in BD designs if the "small bias" condition is not met. The paper confirms that Robust Bias Correction is necessary for valid hypothesis testing and confidence interval construction.
• Heterogeneity is Common: Re-analysis of existing data (e.g., the SPP scholarship study) suggests that treatment effects often vary significantly along the boundary (e.g., varying by wealth or academic performance), which pooling methods average out and obscure.

5. Significance and Recommendations

Significance:
This paper bridges the gap between the practical application of BD designs and rigorous econometric theory. It moves the field from ad-hoc "distance-based" heuristics to a formal framework grounded in geometric measure theory. It validates the use of BD designs for complex, real-world policy evaluations (e.g., school districts, political borders, tax zones) while providing the tools to ensure results are statistically sound.

Recommendations for Practice:

1. For Pooling Studies:
   • Use local regression specifications that include the interaction term between the treatment indicator and the score polynomial.
   • Select bandwidths using MSE-optimal criteria.
   • Conduct inference using Robust Bias Correction (e.g., using rdrobust or rdhte software).
2. For Heterogeneity Analysis:
   • Researchers should move beyond a single average effect and estimate the BATEC.
   • Location-based methods (using bivariate local regression) are recommended as the most robust and general approach, as they handle complex boundary geometries better than distance-based methods.
3. Software Implementation:
   • The authors provide and recommend the use of the R packages rdrobust, rdhte, rdmulti, and rd2d (available at rdpackages.github.io) which implement these theoretically sound methods.

In summary, the paper provides the "gold standard" methodology for analyzing causal effects in settings where treatment assignment is determined by a boundary curve in a multidimensional space, ensuring that empirical findings are both theoretically grounded and practically robust.