Regression Discontinuity Design RDD Checklist: rdrobust Bandwidth, Manipulation Tests & RD Plots

Reviewers reject RDD papers that rely on ad-hoc global polynomials or skip falsification tests. Here is the modern econometric checklist for local linear estimation, data-driven bandwidth selection, density bunching checks, and reproducible plots.

HumanPen Team

· 11 min read

The Modern RDD Falsification and Estimation Protocol at a Glance

A modern regression discontinuity design (RDD) requires four non-negotiable empirical steps: fitting local linear regressions with mean squared error (MSE) optimal bandwidths in `rdrobust`, reporting robust bias-corrected confidence intervals, verifying running-variable density continuity with `rddensity` to rule out cutoff sorting, and testing predetermined baseline covariates and placebo thresholds. Submitting global high-order polynomials or uncorrected ad-hoc bandwidths invites immediate reviewer rejection.

Reviewer standards have shifted permanently away from arbitrary modeling choices: leading econometricians such as Gelman and Imbens, alongside Cattaneo, Idrobo, and Titiunik, have established that credible causal inference in threshold designs hinges on local polynomial approximation near the cutoff rather than flexible curve-fitting across the entire support.
  • Visual inspection and binned RD plots: Construct data-driven binned scatter plots using mimicking-variance or optimal IMSE spacing (such as Stata or R `rdplot`) to verify whether a distinct discontinuity exists at the threshold without imposing parametric assumptions.
  • Local linear estimation and optimal bandwidth selection: Estimate the treatment jump using local linear polynomials (`p = 1`) with triangular kernel weights within MSE-optimal bandwidths (`mserd` or `msetwo`), reserving local quadratic polynomials (`p = 2`) solely for sensitivity checks.
  • Robust bias-corrected inference: Replace conventional confidence intervals with Calonico-Cattaneo-Titiunik robust bias-corrected (RBC) confidence intervals and standard errors to correct for asymptotic bias inherent in data-driven bandwidth optimization.
  • Formal falsification battery: Execute manipulation tests on running-variable density via `rddensity`, demonstrate continuity of predetermined baseline covariates, test artificial placebo cutoffs away from the threshold, and perform donut-hole regressions excluding observations immediately adjacent to the cutoff.

Sharp vs Fuzzy Designs and the Fatal Trap of Global High-Order Polynomials

In a sharp regression discontinuity design, treatment assignment switches deterministically from 0 to 1 at cutoff `c`: every observation with running variable `X >= c` receives the program, and every observation below is excluded. Here, the discontinuity directly identifies the average treatment effect at the threshold. In a fuzzy regression discontinuity design, crossing `c` changes only the probability or intensity of treatment assignment. Fuzzy RDD operates as a local instrumental variables estimator, where the threshold indicator acts as an instrument for treatment take-up, estimating the local average treatment effect for compliers through the ratio of the jump in the outcome to the jump in treatment probability.

For decades, empirical studies routinely fitted global 3rd-, 4th-, or 5th-order polynomials of the running variable across the entire sample support, believing that flexible curves adequately captured nonlinear trends. Gelman and Imbens (2019) dismantled this practice by proving that global high-order polynomials behave erratically: they assign noisy, disproportionately heavy weights to observations far from the cutoff, exhibit Runge's phenomenon with severe boundary oscillations, and produce artificially narrow confidence intervals that dramatically understate inferential uncertainty.

Econometric consensus now strictly favors local low-order polynomials: estimation must be confined to an explicit local neighborhood `[c - h, c + h]` around the threshold, utilizing local linear regression (`p = 1`) with a triangular kernel as the primary specification, and local quadratic regression (`p = 2`) purely to evaluate curvature sensitivity.
  • Extreme and noisy boundary weighting: Global polynomial estimators implicitly weight each observation; high-degree polynomials assign substantial positive and negative weights to points furthest from the cutoff, making boundary treatment estimates volatile and fragile.
  • Severe sensitivity to polynomial degree: Changing the polynomial degree from cubic to quartic or quintic frequently flips the treatment effect sign or alters statistical significance without any substantive economic justification.
  • Poor counterfactual extrapolation: Polynomials of degree 3 or higher curve aggressively near the boundary support, creating spurious gaps at the threshold that mimic genuine treatment discontinuities even when underlying relationships are completely smooth.

Bandwidth Selection Trade-Offs and Robust Bias-Corrected Inference in rdrobust

In local polynomial RDD, selecting the bandwidth `h` represents a fundamental bias-variance trade-off. A wider bandwidth includes more observations, decreasing sampling variance but introducing substantial asymptotic smoothing bias because points further from cutoff `c` deviate from the local linear tangent. A narrower bandwidth reduces approximation bias but inflates sampling variance due to sparse data. To eliminate ad-hoc researcher discretion, the modern standard utilizes data-driven mean squared error (MSE) optimal bandwidth selectors implemented in `rdrobust`, specifically `mserd` for common symmetric bandwidths and `msetwo` for independent bandwidths on each side of the threshold.

Many researchers fall into the trap of constructing conventional confidence intervals around the MSE-optimal point estimate. However, Cattaneo, Idrobo, and Titiunik (2019) demonstrate that the optimal bandwidth `h_MSE` converges at a rate that balances squared bias and variance, which deliberately leaves a non-negligible first-order asymptotic bias in the limiting distribution. As a consequence, conventional t-statistics are not centered at zero under the null hypothesis, causing standard 95% confidence intervals to achieve true empirical coverage rates as low as 80% to 85%.

The Calonico-Cattaneo-Titiunik robust bias-corrected (RBC) method resolves this failure in two integrated steps: it analytically estimates and subtracts the leading bias term using a higher-order polynomial (`p = 2`) over a pilot bandwidth `b`, and then inflates the standard error to explicitly account for the sampling variability of the bias estimator itself, restoring exact asymptotic 95% coverage.
  • Report both conventional and robust estimates: Present the conventional point estimate alongside the robust bias-corrected confidence interval and p-value produced by `rdrobust`. Journal reviewers expect to see both rows in regression tables.
  • Document effective sample sizes: Explicitly disclose the total observations in the dataset alongside the effective sample sizes inside the bandwidth on each side of the threshold (`N_l` and `N_r`), ensuring readers can evaluate localized empirical power.
  • Specify kernel weighting functions: The triangular kernel is mathematically optimal for boundary point estimation under MSE criteria; contrast findings against a uniform or Epanechnikov kernel in supplementary sensitivity checks.
  • State bandwidth choice: Clarify whether the primary specification relies on a common MSE-optimal bandwidth (`mserd`), separate bandwidths (`msetwo`), or coverage-error-rate optimal bandwidths (`cerrd`) for confidence interval optimization.

The Mandatory Falsification Battery: Manipulation, Covariates, and Sensitivity

The causal credibility of regression discontinuity designs rests on the continuity assumption: in the absence of the treatment, potential outcomes must evolve smoothly across threshold `c`. This assumption is satisfied when agents lack the ability or incentive to precisely sort or manipulate their position around the cutoff. If individuals can manipulate their running variable (for instance, students strategically retaking exams to land just above a scholarship score, or firms adjusting revenues right below a tax threshold), sorting generates systematic self-selection. Reviewers therefore demand a comprehensive falsification battery before accepting any treatment effect estimate as causal.

Diagnostic CheckTarget Null HypothesisPrimary Tool / CommandPass CriterionReviewer Action on Failure
Running-variable density manipulationDensity of running variable is continuous at cutoff crddensity (Cattaneo, Jansson & Ma 2018)Robust test p > 0.05, no visual bunchingRejection; design compromised by selective sorting
Predetermined covariate continuityDiscontinuity in baseline covariates equals zero at crdrobust with covariates as pseudo-outcomesRobust RBC p > 0.05 for all pre-treatment traitsDemands explanation; possible confounding policy
Donut-hole sensitivity analysisTreatment effect remains stable after dropping borderline unitsRe-estimate rdrobust excluding [c - delta, c + delta]Point estimate and sign stable as delta expandsDoubt over whether effect is driven by borderline manipulation
Placebo cutoff testsDiscontinuity at artificial thresholds c_placebo equals zerordrobust evaluated at fake thresholds (e.g., medians)No statistically significant jumps (p > 0.05)Rejection; underlying functional form misspecified
Bandwidth sensitivity spectrumEstimates are robust across window widthsRe-estimate across 0.5h, 0.75h, 1.5h, and 2.0hConsistent coefficient signs and overlapping CIsSuspicions of p-hacking via selective bandwidth reporting

The historical standard for testing sorting was the McCrary (2008) density test, which relies on undersmoothed histogram binning followed by local linear smoothing. However, McCrary's test is sensitive to arbitrary choices of bin size and bandwidth. Cattaneo, Jansson, and Ma (2018) introduced `rddensity`, which computes a fully data-driven local polynomial density estimator without requiring pre-binned histograms. It utilizes robust bias correction to provide exact finite-sample inference, reporting both a conventional test statistic and a robust bias-corrected p-value. In empirical write-ups, researchers should report the `rddensity` output along with a continuous density plot illustrating the confidence bands on both sides of cutoff `c`.

  • Baseline covariate balance: Treat predetermined characteristics (such as pre-treatment income, age, gender, or parental education) as pseudo-dependent variables in `rdrobust`. Every single predetermined variable must fail to reject the null hypothesis of zero jump at the threshold.
  • Donut-hole robustness check: Systematically remove observations located in tiny symmetric neighborhoods around the cutoff (such as `[c - 0.01, c + 0.01]` or `[c - 0.02, c + 0.02]`). If point estimates remain stable, you prove that borderline clustering or imprecise scoring does not drive your headline findings.
  • Placebo cutoff validation: Re-estimate the local linear model at artificial cutoff values placed at the 25th, 50th, or 75th percentiles of the running variable where no intervention took place. Statistically significant jumps at fake cutoffs indicate misspecification of the underlying trend.
  • Bandwidth multiplier grid: Generate a bandwidth sensitivity plot displaying point estimates and robust confidence intervals across multipliers ranging from `0.5h` to `2.0h`. Demonstrating stability across this continuum refutes accusations of opportunistic bandwidth tuning.

Automating Binned RD Plots, Falsification Tables, and Scripts in HumanPen Start Project

Turning raw observational records into a complete, publication-ready RDD manuscript section involves substantial computational overhead. Applied researchers must clean running variables, verify cutoff alignments, generate optimal IMSE binned scatter plots via `rdplot`, loop through multi-covariate balance regressions, calculate `rddensity` statistics, and compile multi-column regression tables matching strict journal formatting standards. Managing disparate scripts across Stata, R, and Python often leads to broken cross-checks, mismatched sample counts, and version discrepancies during stressful revision cycles.

Through HumanPen Start Project, researchers can automate these repetitive empirical routines while retaining complete control over theoretical specifications. Attach your data and papers, describe what you need, approve the plan before work starts, and get the reviewed report, charts, and code ready to preview online or download. You can upload CSV or Excel microdata files up to 2 GB alongside your manuscript draft, specify your running variable, cutoff threshold `c`, outcome measures, and candidate baseline covariates, and receive a fully structured econometrics package containing clear estimation tables, automated `rddensity` outputs, and vector-graphic RD plots.

Computational isolation and rigorous privacy govern the execution environment: all empirical estimations run inside an isolated sandbox using clean Python or R code; user data is never shared across accounts or used for model training; projects are retained securely for 6 months (180 days) with anytime deletion. Before submitting to peer review, researchers should also consult our methods and results revision checklist and pre-submission manuscript checklist to ensure all empirical statements align with underlying data.

Every delivered project includes standalone, fully documented replication scripts (such as `01_clean_data.R`, `02_rdrobust_main.R`, and `03_falsification_suite.R`) alongside comprehensive analysis tables. Deliverables serve strictly as an interactive learning scaffold and reproducible methodological reference. The researcher, principal investigator, or graduate author must critically inspect all code routines, evaluate the economic credibility of selected bandwidths against institutional context, and maintain full responsibility for scientific interpretation and institutional disclosure in compliance with journal guidelines.

Frequently Asked Questions About RDD Implementation and Reviewer Critiques

When should I use a local quadratic polynomial (`p = 2`) instead of a local linear model (`p = 1`)? Local linear regression (`p = 1`) should be your primary baseline specification because it minimizes asymptotic boundary bias and avoids overfitting. Local quadratic polynomials (`p = 2`) should be reported alongside local linear models as a secondary sensitivity check. If the estimated treatment effect changes drastically between `p = 1` and `p = 2` within the same bandwidth, it suggests high local curvature near the threshold, requiring careful inspection of your scatter plot.

What should I do if rddensity rejects the null hypothesis of continuity at the cutoff? If `rddensity` yields `p < 0.05`, you cannot simply ignore it or claim that sorting is benign. First, investigate institutional rules to identify why clustering occurs (for example, administrative round-off errors versus deliberate strategic manipulation). Second, perform a donut-hole RDD by dropping observations in a narrow interval `[c - delta, c + delta]` around the cutoff; if the treatment effect remains significant and stable across varying delta values, you provide compelling empirical evidence that borderline manipulators do not drive the main findings.

How should I handle a discrete running variable, such as integer test scores or age in quarters? When the running variable is discrete, the identification assumption of continuous density cannot hold strictly, and observations at the cutoff cannot be shrunk asymptotically. In this setting, clustering standard errors at the values of the running variable (following Lee and Card) was historical practice, but modern guidelines by Cattaneo, Idrobo, and Titiunik recommend local polynomial methods with honest confidence intervals (Kolesár and Rothe) or treating the design as local randomization within a window where units are exchangeable.

What is the practical difference between mserd and cerrd bandwidths in rdrobust? The `mserd` bandwidth selector minimizes the asymptotic mean squared error of the point estimator, striking an optimal balance between point-estimate bias and variance. However, because `mserd` leaves a first-order bias that requires robust bias correction for inference, Calonico, Cattaneo, and Farrell developed the coverage-error-rate optimal bandwidth (`cerrd`), which is asymptotically smaller than `mserd` and specifically minimizes confidence interval coverage errors. Reporting both demonstrates comprehensive sensitivity.

Sources

KEEP READING