Regression Discontinuity

The one idea: Wherever a cutoff decides who gets something — orders over $50 ship free, leads scoring 80+ go to a sales rep — the people just below and just above the line are nearly identical, except one group got treated. Comparing outcomes right at the line turns that ordinary business rule into a free experiment.

⚠️ The trap: "The people just above and just below the cutoff are basically identical, so the jump at the line is the treatment's effect — for all my customers." That feels right and it's wrong. The jump is causal only if nothing else flips at that same line and nobody can game their way across it — and even then it measures the effect only for customers right at the cutoff, not your average or high-value customer.

Somewhere in your product right now, a threshold is assigning treatment. Orders over $50 get free shipping. Leads scoring 80+ get routed to a sales rep. Customers who hit 10,000 points get Gold status. Nobody randomized any of it — yet each cutoff is quietly running a natural experiment that almost nobody bothers to analyze.

Here's the trick, and it's the most intuitive one in causal inference: the customer who scored 79 on your lead model and the customer who scored 80 are, for all practical purposes, the same customer — a coin flip's worth of noise pushed one over the line, and only that one got a sales call. Compare outcomes for people just above and just below the cutoff, and any jump at the line is the effect of the treatment. That's regression discontinuity design (RDD), and Scott…