Testing Density Discontinuity Using Spacings
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Testing Density Discontinuity Using Spacings".
Warren: Here's the gist of the paper. The paper proposes a novel test for assessing density discontinuity in regression discontinuity designs by utilizing spacings between ordered observations to provide an asymptotically valid and robust method for testing continuity assumptions The test offers several advantages over competitors, including applicability, robustness, and simplicity.
Ruth: The Proposed Test Mechanism The core of the proposed test involves using spacings defined as the widths between ordered observations on each side of the cutoff The authors formalize this by assuming that spacings, defined as Di:= X(i) − X(i−1), are approximately independent exponential random variables near the cutoff.
Warren: This approximation suggests using a two-sample test of parameter equality for exponential distributions.
Ruth: The test statistic, Tn, is defined based on these spacings and is designed to be asymptotically unbiased under one framework and consistent under another.
Paper summary: Ruth: So to recap, this paper introduces Testing Density Discontinuity Using Spacings as a novel test for assessing density discontinuity in regression discontinuity designs.
Warren: The central thesis is that by using spacings between ordered observations on each side of the cutoff, they can provide an asymptotically valid and robust method for testing continuity assumptions.
Iris: Essentially, the test claims it has several advantages over competitors, specifically regarding applicability, robustness, and simplicity in this area of econometrics.
Ruth: It’s built around the idea that these spacings yield a continuous statistic that behaves well even under weak assumptions about the underlying density function near the cutoff.
Warren: They study this test across two asymptotic frameworks: one where q is fixed as the sample size increases, and another where q slowly diverges to infinity.
Iris: When q is fixed, they claim exact asymptotic size control without needing any user randomization to achieve it. But when q grows slowly, it becomes consistent without sacrificing that exact size control.
Ruth: The assumptions they require are that the density function has one sided limits at the cutoff and satisfies a Lipschitz condition near that cutoff point.
Warren: This means they avoid needing any derivatives of the density function, which is a significant simplification when you are working with real-world data.
Iris: They compare this approach to things like McCrary's test, which requires three bounded and continuous derivatives of the density on each side of the cutoff.
Ruth: The paper argues that assuming only one sided continuity imposes the weakest requirements on the density function in this context.
Conclusion: Ruth: So, looking at Testing Density Discontinuity Using Spacings by Firmin, it seems the main implication is that there's a simpler test for checking density continuity than we've been using for a long time.
Warren: It offers better control over the tuning parameter and avoids user randomization, which makes it much more practical when you are running these kinds of analyses on actual data.
Iris: What this means for us is that if someone is running an RDD analysis and they suspect their continuity assumption might be wrong, this test gives them a more reliable way to check that suspicion.
Ruth: It suggests that we can test the second implication mentioned in the literature—the continuity of the running variable's density at the cutoff—more reliably with this new approach.
Warren: And overall, it seems like this method is asymptotically efficient among all tests that use a fixed number of observations on each side of the cutoff point.
Iris: It’s a tool that provides higher limiting power in many cases under the fixed q framework, which translates to better detection capability without causing over-rejection under the null hypothesis.
Northwestern University
econ.EM
Submitted: 2026-11-12
Updated: 2026-11-12
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: The paper proposes a novel test for assessing density discontinuity in regression discontinuity designs by utilizing spacings between ordered observations to provide an asymptotically valid and…
Key concepts
- Spacings (Di)
- These are the widths calculated between consecutive ordered observations on either side of the cutoff point in a regression discontinuity design. The test assumes these spacings behave like independent exponential random variables near the cutoff, which forms the basis for testing continuity.
- Asymptotic Frameworks
- The study examines two ways to analyze the test's validity: when the number of observations on each side (q) is fixed, and when q slowly increases toward infinity. The test maintains exact size control in the fixed-q case and consistency as q grows slowly.
- Limiting Distribution (Hθ)
- The paper derives the limiting distribution of normalized spacings by using density transform methods. This distribution depends on the limits of the density function at the cutoff point, denoted as theta, which are defined by $f^+(c)$ and $f^-(c)$. This helps establish how the test statistic behaves asymptotically.
Terminology
Summary
The paper proposes a novel test for assessing density discontinuity in regression discontinuity designs by utilizing spacings between ordered observations to provide an asymptotically valid and robust method for testing continuity assumptions The test offers several advantages over competitors, including applicability, robustness, and simplicity. p. 1 p. 1
The Proposed Test Mechanism
The core of the proposed test involves using spacings defined as the widths between ordered observations on each side of the cutoff The authors formalize this by assuming that spacings, defined as Di:= X(i) − X(i−1), are approximately independent exponential random variables near the cutoff. This approximation suggests using a two-sample test of parameter equality for exponential distributions. p. 4 The test statistic, Tn, is defined based on these spacings and is designed to be asymptotically unbiased under one framework and consistent under another. p. 1
Asymptotic Frameworks and Validity
The test is studied in two asymptotic frameworks: one where q is fixed as the sample size increases and one where q slowly diverges to infinity. When q is fixed, the test achieves exact asymptotic size control without user randomization. p. 0 When q increases slowly, the test becomes consistent without sacrificing exact asymptotic size. p. 0 The assumptions required for validity include that the density function f(x) has one-sided limits at the cutoff c and satisfies a Lipschitz condition near c. p. 2
Performance Against Alternatives
The spacings test is compared against alternatives like McCrary's test, which requires three bounded and continuous derivatives of the density on each side of the cutoff. The authors argue that assuming only one-sided continuity imposes the weakest requirements on the density. p. 2 Furthermore, in Monte Carlo simulations, the spacings test controls size better than Bugni and Canay's (2021) g-order statistics test in most designs. The comparison shows that the spacings test avoids restrictions on the tuning parameter and user randomization. p. 1
Practical Implementation and Tuning
The test requires a choice of the number of local observations, q, which can be determined using a rule-of-thumb based on asymptotic bias bounds. The authors propose a rule-of-thumb formula to choose q that optimizes size control by estimating parameters like f(c) and κ using a normal distribution approximation. p. 7 This method allows researchers to focus on the trade-off between bias and power without being overly sensitive to the precise choice of q. The computational details involve a bisection search for critical values based on solving an expected value equation. p. 5
Conclusion on Efficiency
The paper concludes that the spacings test is asymptotically efficient among tests that use q observations on each side of the cutoff in the fixed-q framework. When considering q to slowly increase to infinity, the spacings test and g-order statistics test possess similar limiting properties. p. 0 The test's ability to control size precisely, regardless of q, makes it a superior choice over tests that rely on discrete statistics or restrictive tuning parameters. This suggests the spacings test is a valuable tool for testing density discontinuity in RDD. p. 0
How the training signal is built
The limiting distribution of the normalized spacings, Hθ, is derived by expanding the spacing statistics and using density transform methods to find its probability density function. This limiting distribution depends on the limits of the densities at c, denoted as θ = (f+(c), f−(c)). p. 7 p. 7 The test statistic Tn converges in distribution to a statistic from the two-sample exponential problem of Han et al. p. 5 p. 5 (2012) under certain conditions. p. 5 The asymptotic normality of the test statistic is established through central limit theorems for a random number of summands. p. 8
Where the gains come from
The primary gain of the spacings test over Bugni and Canay's (2021) test lies in its ability to provide exact asymptotic size control without user randomization or restrictions on q. The continuous nature of the test statistic avoids the sensitivity to tuning parameter choices that plagues other tests. p. 1 p. 1 In Monte Carlo simulations, the spacings test generally controls size better than both Bugni and Canay's test and Cattaneo et al.'s (2020) local polynomial test. The paper demonstrates that the spacings test provides uniformly more limiting power in many cases under the fixed-q framework. p. 3 p. 3 This superior performance is attributed to its exact asymptotic size control, which translates into higher limiting power without causing over-rejection under the null. p. 2 p. 2 The results hold even when considering q to slowly increase to infinity, where the spacings test and g-order statistics test possess similar limiting properties. p. 0 p. 0 The paper demonstrates that any differences disappear as Q increases due to the continuity of the test statistic. p. 1 p. 1 The final conclusion is that the spacings test is asymptotically efficient among tests that use q local observations on each side of the cutoff. p. 2 p. 2
Pruning redundant reasoning tokens
The paper proposes a new method to test for density discontinuity in RDD by using spacings between ordered observations to provide an asymptotically valid and robust method for testing continuity assumptions The test offers several advantages over competitors, including applicability, robustness, and simplicity.
The primary gain of the spacings test over Bugni and Canay's (2021) test lies in its ability to provide exact asymptotic size control without user randomization or restrictions on q. The continuous nature of the test statistic avoids the sensitivity to tuning parameter choices that plagues other tests. In Monte Carlo simulations, the spacings test generally controls size better than both Bugni and Canay's test and Cattaneo et al.'s (2020) local polynomial test. The paper demonstrates that the spacings test provides uniformly more limiting power in many cases under the fixed-q framework. This superior performance is attributed to its exact asymptotic size control, which translates into higher limiting power without causing over-rejection under the null. The results hold even when considering q to slowly increase to infinity, where the spacings test and g-order statistics test possess similar limiting properties. The paper demonstrates that any differences disappear as Q increases due to the continuity of the test statistic. The final conclusion is that the spacings test is asymptotically efficient among tests that use q local observations on each side of the cutoff. The paper proposes a new method to test for density discontinuity in RDD by using spacings between ordered observations to provide an asymptotically valid and robust method for testing continuity assumptions. p. 1 The test offers several advantages over competitors, including applicability, robustness, and simplicity.
The limiting distribution of the normalized spacings, Hθ, is derived by expanding the spacing statistics and using density transform methods to find its probability density function. This limiting distribution depends on the limits of the densities at c, denoted as θ = (f+(c), f−(c)). The test statistic Tn converges in distribution to a statistic from the two-sample exponential problem of Han et al.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved density discontinuity testing by implementing a novel test using q spacings on each side of a cutoff, which yields a continuous statistic behaving well under weak assumptions and achieving exact asymptotic size control without user randomization.
-
Enhanced robustness against tuning parameter choices by utilizing a continuous test statistic that avoids the need for user randomization and simplifies methods for choosing the tuning parameter, performing similarly to Bugni and Canay’s (2021) test when their tuning parameter is well-chosen.
-
Increased applicability across different asymptotic frameworks by providing asymptotically valid inference under both a fixed q framework (as sample size n increases while q remains fixed) and an increasing-q framework (as q slowly diverges to infinity).
-
More efficient testing in the increasing-q framework by achieving consistency when q increases slowly, which is required for finitesample applications where researchers are comfortable using enough local observations.
-
Data-driven tuning parameter selection via a rule of thumb that provides an initial guess for the Lipschitz constant and density limit using a normal distribution approximation, allowing researchers to specify an allowable amount of bias t star and solve for q star.
-
Superior performance in power comparison against alternatives by providing uniformly more limiting power in many cases under the fixed-q asymptotic framework compared to the g-order statistics test of Bugni and Canay (2021).
-
Enhanced finite sample size control by demonstrating that the spacings test controls size better than tests like BC, particularly in designs where severe nonnormality is suspected, by allowing for a more cautious choice of q.
-
Better handling of density discontinuities under difficult scenarios by showing that when density limits are close but not equal, the test can detect them with probability exceeding alpha by a certain margin determined by the ratio of density limits.
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs