Free explainer3 min read
Stratified randomisation
Understand the model, recognise it in a stem, separate the look-alikes, then apply it.
Start with the mental model
Why isn't it enough to simply flip a coin for every trial participant? Simple randomisation (allocating each participant to a group purely by chance, with no adjustment for anything) protects against selection biasA systematic error from how participants were chosen, making the study sample unrepresentative.Read more → on average, but in a small trial, chance alone can still leave one arm with far more of a key prognostic factor than the other — a factor known, before the trial starts, to affect the outcome (such as baseline disease severity). Stratified randomisation fixes this by pre-specifying that factor and guaranteeing balance on it directly, rather than hoping randomness delivers balance by luck. For example, imagine a trial of a new antidepressant with only 40 participants, where baseline symptom severity strongly predicts response regardless of treatment; if chance allocation happened to put most of the severe cases in one arm, that arm's results would be confounded by severity, not just by the drug. In practice, this means the researchers instead split recruits into a mild-severity stratum and a severe-severity stratum in advance, and randomise separately within each one.
01Core model
One or more prognostic factors (patient characteristics known or strongly suspected, before recruitment begins, to affect the outcome — for example, age band, baseline disease severity, or recruiting site) are chosen in advance and used to divide potential participants into strata (subgroups), each defined by one combination of those factors. As each participant is recruited, they are assigned to their matching stratum, and randomisation — usually block randomisation, which allocates participants to arms in balanced, small batches called blocks — is then carried out separately within each stratum. Because each stratum is randomised on its own, the proportion of participants from that stratum ending up in each treatment arm stays balanced throughout recruitment, not just on average once the trial is complete. This differs from simply randomising the whole sample once with no strata: that method balances the arms only in the long run and offers no protection against imbalance on any one specific factor, especially when the total sample is small. Stratified randomisation is most useful when there are few strata (one or two prognostic factors); as the number of stratifying factors grows, the number of strata multiplies quickly, and some strata can end up with too few participants for the balancing to work well.