Randomisation and Stratification

Randomisation allocates each participant to an arm by a chance mechanism that nobody involved can predict or influence. Its purpose is to make the arms exchangeable in expectation on every prognostic factor, measured or not, the property no observational adjustment can reproduce. Simple randomisation leaves noticeable imbalance in small trials, so block randomisation with permuted, undisclosed block sizes is used to keep arm sizes close without making the next assignment guessable.

Stratified randomisation runs a separate allocation sequence within each level of a factor known to drive outcome. Multicentre trials nearly always stratify by site, and metabolic trials by baseline glycated haemoglobin, body mass index category or background therapy. Minimisation is the adaptive alternative, assigning each new participant to whichever arm reduces the current imbalance.

Randomisation is what licenses causal language. Without it, those who received the drug differ from those who did not by the very reasons they were selected, and confounding by indication cannot be fixed afterwards. What randomisation guarantees is balance on average rather than balance in any particular trial, which is why a baseline table with a chance imbalance is expected and why significance testing of baseline characteristics is meaningless.

The word gets applied to things that are not random: alternate allocation, day of the week, hospital number or date of birth, clinician judgement. All of these are predictable in advance, which means the person recruiting can steer who enters which arm, and trials using them show systematically larger effects. The other failure is stratifying on too many variables in a small trial, which fragments the sequence and defeats the balance it was meant to create.