Randomisation Methods

Randomisation is how the allocation sequence is generated. The common methods are: simple (each patient allocated independently, like a coin toss); block (randomise within small blocks so arms stay balanced in size); stratified (randomise within strata of key prognostic factors so those factors are balanced across arms); and minimisation (an adaptive method that allocates each new patient to whichever arm keeps several prognostic factors most balanced). The aim is comparable groups plus a sequence no one can predict.

Two separate steps. Generating the sequence is randomisation proper; the method chosen (simple, block, stratified, minimisation) controls how well-balanced the arms are. Concealing the sequence — allocation concealment — is the separate guarantee that whoever enrols a patient cannot foresee or steer the next allocation. A trial needs both.

Ask first “how was the sequence made?” Simple randomisation can leave arms unequal in small trials; block randomisation fixes group sizes; stratification and minimisation balance prognostic factors. Then ask the different question “could the recruiter predict the next allocation?” — that is concealment. Methods tied to a predictable rule (alternate days, odd/even hospital number, day of the week) fail the first test: they are quasi-random and can be gamed.

Good sequence generation produces comparable groups, removing confounding by indication — the whole point of an RCT. A predictable or subvertible sequence lets sicker (or healthier) patients be channelled into one arm, reintroducing selection bias and inflating or masking the treatment effect.

• Simple = independent coin-toss; can imbalance small trials
• Block = balances arm sizes · Stratified = balances key factors
• Minimisation = adaptively balances multiple factors
• Generation ≠ allocation concealment (hiding the sequence)

CRASH-2 (Lancet 2010) — 20,211 bleeding trauma patients, TXA vs placebo. The sequence was computer-generated with a block size of eight, balanced by centre. Most patients were randomised by opening the next consecutively numbered, identical treatment pack; sites with reliable phone/internet access could instead use the central Oxford randomisation service (a minimisation algorithm) — only 95 of the 20,211 went this route. The identical, pre-numbered packs (and indistinguishable TXA vs placebo ampoules) gave robust allocation concealment and blinding — the recruiter could not predict or influence the next allocation.

• Treating quasi-randomisation (alternate days, DOB, hospital number) as truly random.
• Forgetting predictable block sizes can be guessed in an unblinded trial — vary or mask block size.
• Confusing sequence generation with allocation concealment — they are different safeguards.