Randomisation

Who gets the treatment? If you choose, you might, without meaning to, give it to the healthier patients, the keener students, or the easier test cases. Then any difference could come from your choice, not the treatment.

Randomisation means letting chance decide who goes into which group.

Why a coin beats your judgement

• It balances the things you know about, like age or skill.
• It also balances the things you don't know about, or can't measure.
• It removes the experimenter's own bias in choosing.

The first modern trial

In 1948, the UK Medical Research Council published a trial of the antibiotic streptomycin for tuberculosis. Patients were assigned to treatment or control using random numbers in sealed envelopes. It is widely seen as the first modern randomised controlled trial, and set the pattern for medicine since.

How to randomise

• Use a random number generator, not "every other person" or "whoever arrives first".
• Hide the assignment until the unit is enrolled, so it can't be influenced.
• In ML, randomise which data goes into train and test, and which random seeds are used.

When groups are small

With few units, chance can still create imbalance. Blocking helps: group similar units (same school, same hardware) and randomise within each block.

RememberLet chance assign the treatment. It balances what you know, and what you don't.