The bandwidth in a [[Design Pattern II - Regression Discontinuity (RD)|regression discontinuity]] design controls a bias-variance tradeoff. A wider bandwidth uses more observations (lower variance) but includes units farther from the cutoff where the local polynomial approximation is less accurate (higher bias). A narrower bandwidth keeps only the most comparable units (lower bias) but with fewer observations (higher variance). The **MSE-optimal** bandwidth minimizes the mean squared error of the estimator, $\text{MSE}(h) = \text{Bias}(h)^2 + \text{Variance}(h)$. An alternative is the **CER-optimal** (coverage error rate) bandwidth, which minimizes the coverage error of the confidence interval rather than the MSE of the point estimate. It is typically narrower than the MSE-optimal bandwidth. > [!info]- Last updated: April 12, 2026