Regression discontinuity design (RDD) is a quasi-experimental causal inference method that identifies the effect of a treatment by exploiting a known cutoff rule for assignment, where units just above and just below the threshold are comparable except for treatment status.
RDD provides highly credible causal estimates from observational data where randomized controlled trials are infeasible, directly informing high-stakes policy and business decisions. This rigor translates into optimized resource allocation, reduced risk, and defensible strategies, directly impacting ROI and regulatory compliance.
1Careers
1Categories
8.7 Avg Demand
15%
Avg AI Risk
How to Learn Regression discontinuity design (sharp and fuzzy)
1. Understand the core logic: The cutoff creates a local randomized experiment. 2. Grasp the distinction between Sharp RDD (treatment assignment is a deterministic function of the running variable) and Fuzzy RDD (treatment assignment is probabilistic, used as an instrument). 3. Master the graphical analysis: plotting binned averages of the outcome against the running variable to visualize the discontinuity.
1. Move to formal estimation: implementing local linear regressions with optimal bandwidth selection (e.g., using Imbens-Kalyanaraman or Calonico-Cattaneo-Titiunik methods). 2. Conduct rigorous validation: testing for continuity of pre-determined covariates at the cutoff (balance checks) and manipulation tests for the running variable (McCrary density test). 3. Avoid common pitfalls: over-reliance on global polynomials, ignoring sensitivity to bandwidth choice, and misinterpreting fuzzy RDD estimates as ITT instead of LATE.
1. Handle complex designs: multi-cutoff RDD, geographic RDD, and regression kink designs. 2. Integrate RDD into broader strategic analysis: combining it with cost-benefit analysis, understanding its external validity limitations (local average treatment effect), and communicating results to non-technical stakeholders. 3. Mentor teams: develop internal RDD review standards and assess the validity of external RDD studies for strategic partnerships or acquisitions.
Practice Projects
Beginner
Project
Sharp RDD on Scholarship Eligibility
Scenario
A university provides a scholarship to students with a high school GPA above 3.5. Analyze the causal effect of the scholarship on first-year college GPA.
How to Execute
1. Obtain a dataset with high school GPA, scholarship receipt, and college GPA. 2. Center the running variable (GPA) at the cutoff (3.5). 3. Create a scatter plot of college GPA vs. centered high school GPA with binned averages. 4. Run separate local linear regressions on each side of the cutoff to estimate the jump.
Intermediate
Case Study/Exercise
Fuzzy RDD in Corporate Lending
Scenario
A bank uses a credit score threshold (e.g., 700) to *automatically approve* loans, but loan officers have discretion to approve some applicants just below the threshold. Estimate the causal effect of loan approval on default rates.
How to Execute
1. Treat the credit score as the running variable and the 700-point cutoff as the threshold. 2. Use the fuzzy RDD approach: instrument actual loan approval with the *eligibility* indicator (score >= 700). 3. Estimate the first-stage (effect of eligibility on approval) and the reduced-form (effect of eligibility on default) regressions. 4. Compute the Wald estimator (ratio of reduced-form to first-stage) to get the Local Average Treatment Effect (LATE) for compliers.
Advanced
Project
RDD for Regulatory Impact Assessment
Scenario
A new environmental regulation mandates pollution controls for factories exceeding a size threshold (e.g., 500 employees). Estimate the regulation's causal impact on firm productivity and emissions.
How to Execute
1. Define the running variable (employee count) and cutoff (500). 2. Test for manipulation around the cutoff (do firms cluster just below?). 3. Use a multi-cutoff design if thresholds vary by region. 4. Estimate RDD models controlling for industry and regional fixed effects. 5. Conduct cost-benefit analysis: compare the productivity loss (the RDD estimate) against the estimated social benefit of reduced emissions to inform policy advocacy.
These are industry-standard tools for implementing modern RDD methods, including optimal bandwidth selection, bias-corrected estimation, and robust inference. Use them for all serious empirical work.
Methodological Frameworks
Imbens-Kalyanaraman (IK) BandwidthCalonico-Cattaneo-Titiunik (CCT) MethodMcCrary Density Test
The IK and CCT frameworks provide the econometric foundation for optimal bandwidth selection and bias correction. The McCrary test is a mandatory diagnostic to check for sorting manipulation around the cutoff.
Interview Questions
Answer Strategy
Structure the answer around the 4 core RDD steps: 1) Design (sharp RDD, running variable is rating, cutoff is 90), 2) Visualization (plot next-year rating vs. current rating with bins), 3) Estimation (local linear regression on each side), 4) Validation (covariate balance, McCrary test). Emphasize the key assumptions: continuity of potential outcomes at the cutoff and no precise manipulation.
Answer Strategy
Test for methodological rigor and practical judgment. The answer should question the bandwidth choice (too wide may introduce bias from functional form misspecification), suggest using data-driven optimal bandwidth selection (e.g., IK or CCT), and recommend presenting results across a range of bandwidths as a robustness check. Frame it as ensuring the finding is not an artifact of modeling choices.
Careers That Require Regression discontinuity design (sharp and fuzzy)