🎰Randomisation

Deciding by chance which units get the treatment and which do not. Not by convenience, not by who volunteered and not by the experimenter's judgment.

Why it protects a comparison: anything that affects the outcome, including things nobody thought to measure, is spread across both groups in roughly equal amounts. The groups then differ systematically in one way only, the treatment, and a difference in results can be read as its effect. Chance also gives an honest basis for saying how large a difference could arise from the allocation alone.

🏛️ Ronald Fisher made the argument in the 1920s from agricultural field trials, where the obvious method was to give the new fertiliser to the plots that looked suitable.

Example from work: 📌 Design the triage A-B test. The tempting design was to switch auto-triage on for the customers who had asked for it. Those are the keen ones, with tidier data. Any result would have been about keenness.

Often confused with: a random sample, which is about who is in the study at all. Randomisation is about who gets what once they are in.

Taught in 🎓 Lecture 9 - A-B tests and their traps, the most useful evening of the spring.