Regression Discontinuity, explained.
Regression discontinuity uses a treatment assignment threshold to compare outcomes close to the cutoff, assuming other potential-outcome determinants vary smoothly there.
Why it happens
Scores are −10 through −1 and +1 through +10. Outcomes equal 50 + 0.5 times score plus a ten-unit treatment jump above zero. An optional second policy adds five at the same cutoff. The display compares means in symmetric windows, not fitted one-sided limits.
Regression discontinuity uses a treatment assignment threshold to compare outcomes close to the cutoff, assuming other potential-outcome determinants vary smoothly there.
Read the result
The dots show a smooth trend plus a jump. The reported mean contrast includes both; narrowing the window reduces the trend contribution without removing it completely.
A worked example
A scholarship threshold with a second benefit
Use five points on each side of the cutoff. The raw contrast is thirteen despite a treatment jump of ten.
Narrow to one point on each side. The contrast becomes eleven because the two retained scores still differ by two.
Turn on the other policy. The contrast gains another five, which cannot be separated from treatment using this cutoff alone.
OPTIONAL DEEPER DETAILGo deeper: inside the model
Inside this model
Scores are −10 through −1 and +1 through +10. Outcomes equal 50 + 0.5 times score plus a ten-unit treatment jump above zero. An optional second policy adds five at the same cutoff. The display compares means in symmetric windows, not fitted one-sided limits.
Where this idea is useful
A practical use
Eligibility thresholds can help investigate a scholarship programme when applicants cannot precisely manipulate their scores.
A common misconception
“Nearby observations make every cutoff comparison causal.”
Continuity, assignment behaviour and other interventions matter. A coincident policy jump defeats attribution even when scores are very close.
What this explanation leaves out
- This is a raw local contrast with discrete scores, not an RD estimate with uncertainty. A smaller window retains a smooth-trend contribution. Another discontinuous intervention breaks attribution to the target policy alone.
Why does the narrowest window not return exactly ten?
Scores are discrete and the display uses group means rather than fitted limits. The retained −1 and +1 observations still have a one-unit baseline difference.
Could another rule change at exactly the threshold you want to study?
Associated thinkers
Associations marked provisional are awaiting source review.
Further reading
Explore the original research or the teaching reference behind this experiment.