rdd-analysis

Regression discontinuity design estimates local treatment effects by fitting local polynomial models on a running variable around a defined cutoff. You need a running variable, a cutoff point, and outcome data to compute robust in-sample estimates.

Sharp RDD assumes perfect treatment assignment at the cutoff, while fuzzy RDD handles imperfect compliance by using the cutoff as an instrumental variable. Both designs require bandwidth selection and local polynomial estimation to identify causal effects.

Optimal bandwidth selection balances bias and variance near the cutoff to produce robust local treatment estimates. Established RDD tools apply data-driven algorithms to determine the ideal window for local linear estimators.

RDD validity testing includes density tests of the running variable, covariate balance checks, and placebo cutoffs. These tests ensure no manipulation occurred at the threshold and that covariates are continuous across the cutoff.

Yes, a triangular kernel is commonly applied to weight observations closer to the cutoff more heavily during local linear estimation. Specifying this kernel alongside your running variable and outcome produces robust discontinuity estimates.

RDD is inappropriate when no defined cutoff exists in the running variable or when treatment assignment is not determined by a threshold. It specifically estimates local average treatment effects at the cutoff, not global effects across the entire sample.