Audit bandwidth choices in regression discontinuity

Establish the assignment mechanism

Regression discontinuity uses a treatment-assignment threshold in a running variable to study outcomes close to that threshold. Begin with the actual policy rule, including exceptions, rounding and whether treatment receipt changes sharply or only probabilistically at the cutoff. A convenient split in a continuous variable is not enough to create a credible design.

The developers' rdrobust documentation describes local-polynomial estimation, bandwidth selection and robust bias-corrected inference. These tools implement an analysis; they do not establish the institutional assumptions that make the cutoff informative.

Audit observations around the cutoff

Plot the running variable and outcome using transparent binning choices. Examine whether units can manipulate their position near the threshold and whether recorded values are heavily rounded. Review continuity of relevant baseline characteristics and other policies at the same cutoff. Explain which findings would weaken the interpretation rather than treating diagnostic p-values as a certificate of validity.

Define the bandwidth rule before choosing a favourable estimate. The bandwidth controls how local the comparison is and affects both precision and sensitivity to functional form. Preserve the selected bandwidth, kernel, polynomial order and inference method in the analysis record. Avoid fitting a high-degree global polynomial merely to produce a smooth-looking curve.

Hypothetical example

Suppose a fictional programme becomes available at a score of 50. An analysis using scores from 45 to 55 studies a different local window from one using scores from 20 to 80. If the estimated jump changes substantially between defensible local bandwidths, report that instability and investigate support, curvature and data quality.

If some eligible people do not receive the programme, record the change in treatment receipt at the cutoff. A fuzzy design requires additional assumptions and should not be reported as if assignment and receipt were identical.

Interpret the local result

State the population near the cutoff to which the estimate applies. Do not automatically generalise to people far from it or to another threshold. Present sensitivity analyses for justified bandwidths and specifications, while retaining the prespecified main result.

Use the range-rule worksheet to check impossible scores and preserve legitimate boundary values. Archive a synthetic test that verifies how equality to the cutoff is assigned. The final report should connect the numerical estimate to the assignment mechanism, uncertainty and limits of local interpretation.

Source verification: 6 November 2026. Examples are hypothetical.