RDD Bandwidth Selection and Local Estimation
You'll learn how to choose the right window around the threshold and use local polynomial regression to estimate causal effects in Regression Discontinuity Designs.
The Core Challenge
In RDD, you estimate treatment effects by comparing observations just above and just below a threshold. But how close is "just"? The bandwidth is the width of the window on each side of the cutoff that you include in your analysis.
Why Bandwidth Matters
Think of it like choosing binoculars: too narrow (small bandwidth) and you see fine detail but include too few observations, creating noisy estimates. Too wide (large bandwidth) and you include more data but risk capturing observations far from the cutoff where treatment and control groups differ for reasons beyond the threshold.
Local Polynomial Regression
Once you select a bandwidth, you don't just compare raw means. You use local polynomial regression—typically linear or quadratic—to fit separate trend lines on each side of the threshold. This accounts for any underlying relationship between the running variable and outcome, isolating the discontinuous jump at the cutoff as your causal estimate.
Common Bandwidth Selection Methods
- Rule-of-thumb approaches: Use formulas based on sample size and variance
- Cross-validation: Minimize prediction error
- Optimal bandwidth selectors: Methods like Imbens-Kalyanaraman or Calonico-Cattaneo-Titiunik that balance bias and variance
The bandwidth-treatment effect trade-off is fundamental: smaller bandwidths reduce bias from functional form misspecification but increase variance from fewer observations.
Key Takeaway: Bandwidth selection balances bias and precision—use data-driven methods to choose the window around your RDD threshold, then apply local polynomial regression to estimate the causal discontinuity while controlling for smooth trends on either side.