Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Development

rdd-diagnostics

Diagnostics for regression-discontinuity and regression-kink designs — sharp, fuzzy, discrete-score, multi-cutoff and geographic: which output row carries the verdict, bandwidth sensitivity, density and manipulation tests with manipulation-robust bounds, covariate continuity and covariate adjustment, placebo cutoffs, donut checks, RD plots, local-randomization inference, boundary-discontinuity estimation, extrapolation claims, heterogeneity and power. Invoke from `code-and-data` whenever a target result uses a cutoff, a kink in a policy schedule, or a boundary.

isitcredible · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add isitcredible/research_skills --skill rdd-diagnostics

About this skill


name: rdd-diagnostics description: >- Diagnostics for regression-discontinuity and regression-kink designs — sharp, fuzzy, discrete-score, multi-cutoff and geographic: which output row carries the verdict, bandwidth sensitivity, density and manipulation tests with manipulation-robust bounds, covariate continuity and covariate adjustment, placebo cutoffs, donut checks, RD plots, local-randomization inference, boundary-discontinuity estimation, extrapolation claims, heterogeneity and power. Invoke from code-and-data whenever a target result uses a cutoff, a kink in a policy schedule, or a boundary.

RDD diagnostics

Re-run the design the way its method's authors prescribe, then test whether units could sort across the cutoff. The published fit shows the cutoff at its most flattering — one bandwidth, one polynomial, one kernel, one plot. The checks below hold it to the published procedure (robust bias-corrected inference at data-driven bandwidths, the falsification battery) and decide whether the design identifies at all.

Trigger: any target result identified off a cutoff, a kink, or a boundary — sharp, fuzzy, discrete-score, multi-cutoff, or geographic. NOT_APPLICABLE when no meaningful cutoff exists; do not run placebo cutoffs on a design without one.

Method cribs (source-audited; the index and what each governs is ../../references/methods/rdd_methods_guide.md §0). Cite the underlying methods work (author, year, and page or section) behind every verdict-bearing rule you apply, never the crib filename or crib section number. Code is in the battery's ../../references/software-examples.md.

Software versions

Record the installed version of every RD package you run (packageVersion in R, which in Stata) and, where the replication package states it, the version the authors used; check the version after installing, because prebuilt binaries can lag the source release. The packages change numbers and calls between releases, so a mismatch with a paper's printed bandwidth or estimate can be a version artefact rather than an author error:

  • rdrobust 3.0.0: HC2/HC3 bias-regression residuals use the q-order leverage, so robust SEs under vce = "hc2"/"hc3" move slightly.
  • rdrobust 4.0.0: clustered variance remapped — with cluster, the default vce = "nn" silently becomes CR1, hc0/hc1 become CR1, hc2/hc3 become CR2/CR3 — so every clustered bandwidth, SE and p-value changes (by a few percent with dozens of clusters, by multiples with a handful); all/detail moved from rdrobust() to summary(); the legacy CCT/IK/CV selectors (rdbwselect_2014) were removed.
  • rdrobust 4.1.1 (October 2026; Python 2.1.x, Stata 11.1.x): standardisation inside bandwidth selection on by default; padded bwcheck floor; default ginv.tol 1e-15; a halved point estimate in sharp designs with covariates and deriv >= 2 corrected; a stop with "Not enough variability" when a side has too few distinct running-variable values; a warning when either side has fewer than 10 clusters inside the bandwidth. Unclustered default fits agree with earlier releases to about seven significant figures.
  • rdlocrand 3.0: window-selection and mass-point indexing bugs in 1.1 and 2.0 fixed, so selected windows and downstream p-values change; polynomial (p > 0) inference is HC-robust large-sample (HC3 default) and no longer reports a permutation p-value; unsupported polynomial requests fall back to p = 0 with a warning.
  • rddensity 2.5 and 3.0 give the same numbers; both run the mass-point adjustment and a table of binomial tests by default, which the 2018 software article predates.
  • Current rdpower and rdhte require rdrobust ≥ 4.0.0; where a project must keep an older rdrobust to reproduce a paper, install the current packages in a separate library.

Never allege an error in a paper's RD numbers until the version is matched or the gap is shown to survive it. Run the current release for the review's own estimates and say so.

Core checks

  • Which row carries the verdict. High-order global polynomials are not defensible estimates; re-estimate with a local linear fit (triangular kernel) at a data-driven bandwidth. The headline is the conventional point estimate at the MSE-optimal bandwidth with the robust bias-corrected (RBC) interval and p-value; add the RBC interval at the CER-optimal bandwidth. The robust interval is centred on the bias-corrected point; quote it as printed and never re-centre it on the conventional estimate. In rdrobust ≥ 4.0.0 the default summary() prints one row — the conventional point estimate beside the robust z, p and interval; summary(f, all = TRUE) prints the conventional, bias-corrected and robust rows (in ≤ 3.0.0, rdrobust(..., all = TRUE)); f$coef, f$se, f$pv, f$ci hold the three rows in that order in every version. vcov() on an rdrobust fit is diagonal with placeholder zeros — never use it to test one row against another. A conventional SE, interval or star at an MSE- or CER-optimal bandwidth is the combination the selector's authors declare invalid and is reported only as a diagnostic. Four traps (calonico_cattaneo_farrell_rbc_bandwidth.md):
    • a t-ratio formed as the conventional coefficient over the robust SE is neither test. When only the robust interval is printed, its midpoint is the bias-corrected estimate and half its width over 1.96 is the robust SE; a printed z or p inconsistent with them means rows were mixed;
    • a point estimate taken from a cerrd/certwo run is undersmoothed, not MSE-optimal;
    • a local-quadratic interval at the local-linear MSE bandwidth is the RBC interval with ρ = 1 — not an independent robustness check;
    • a bootstrap of the conventional estimator, or an ad hoc smaller bandwidth with conventional SEs, does not remove the bias.
  • **Bandwidth and order sensitivity over a small ra
Read the complete source ↗