Course contentsShow
← Data Science
Lesson 1462 of 2,14536. Causal Inference BasicsFree lesson

Identifying Valid RD Thresholds

Evaluating whether assignment rules create legitimate discontinuities and checking for manipulation around the cutoff.

Identifying Valid RD Thresholds

You'll learn how to evaluate whether a cutoff creates a legitimate discontinuity and check if subjects manipulate their assignment around that threshold.

When using Regression Discontinuity Design, the entire causal claim hinges on one critical assumption: the threshold must be a true discontinuity where assignment changes abruptly, not a point that subjects can manipulate. If people can game the system to land just above or just below the cutoff, your RD design collapses.

What Makes a Threshold Valid?

A valid RD threshold requires:

  1. A clear, enforceable rule – The assignment must follow a strict, predetermined cutoff (e.g., "students scoring ≥70 get tutoring"). Fuzzy boundaries or subjective decisions undermine validity.

  2. No manipulation around the cutoff – Individuals cannot precisely control their assignment variable to land on their preferred side. Think: you can't choose your exact birthdate, but you might retake a test strategically.

Testing for Manipulation

The density test (McCrary test) checks whether unusually many observations cluster just above or just below the threshold. If the density of the assignment variable shows a suspicious jump right at the cutoff, that's a red flag. For example, if students scoring 69.9 are rare but 70.1s are abundant, they might be influencing their scores.

Analogy: Imagine a bridge that closes at exactly 5:00 PM. If cars naturally arrive smoothly throughout the day, the cutoff is valid. But if you see a suspicious pile-up of cars arriving at 4:59 PM (and almost none at 5:01 PM), drivers are manipulating their arrival times—breaking your design.

Why It Matters

Without a valid threshold, treatment and control groups aren't comparable anymore. Manipulation introduces selection bias, making your causal estimates unreliable.

Key Takeaway: A valid RD threshold must be rule-based, enforceable, and free from manipulation—test the density around the cutoff to verify subjects couldn't game the system.