Randomization

The process of assigning trial participants to treatment groups by chance, producing groups comparable on both known and unknown influencing factors.

Source: Fisher 1935

Assigning patients to groups by chance means that, in a large enough trial, every characteristic, whether recorded or not, tends to be spread evenly between the groups. This is randomisation's great strength, since it balances the confounders a researcher knows to measure and, just as importantly, those they have never thought of or cannot measure, which no statistical adjustment could handle. The groups therefore start alike in every respect except the treatment, so a difference in outcome can be attributed to the treatment. Chance balances the unknown. Schulz and Grimes (2002) explain this.

Source: Schulz & Grimes 2002

Randomisation works by using a chance mechanism, such as random number generation, to allocate participants to treatment groups, so that the assignment is not determined by the participant's or investigator's choice. Methods include simple randomisation, block randomisation to balance group sizes, and stratified randomisation to balance important factors. The allocation is concealed until enrolment to prevent bias. Because assignment is by chance, the groups tend to be similar in all characteristics, both measured and unmeasured, so any difference in outcome can be attributed to the treatments rather than to systematic differences between the groups.

Source: Friedman, Furberg & DeMets 2015

Randomisation is important because it balances both known and unknown confounding factors across treatment groups, creating comparable groups so that differences in outcome can be attributed to the treatments rather than to pre-existing differences. This is what allows randomised trials to establish causal effects with less bias than observational studies, where groups may differ systematically. Randomisation also supports valid statistical inference and, with allocation concealment, prevents selection bias in forming the groups. Because it underpins the internal validity of a trial, randomisation is the key feature distinguishing experiments from observational studies.

Source: Fisher 1935

Randomisation controls confounding by allocating participants to groups purely by chance, so that confounding factors, whether measured or unmeasured, known or unknown, tend to be distributed similarly across the groups rather than being associated with the treatment. Because the groups are comparable on these factors on average, they cannot systematically bias the comparison, so the difference in outcomes reflects the treatments. This ability to balance unknown as well as known confounders is unique to randomisation, distinguishing it from analytic adjustment, which can address only measured confounders, and giving randomised trials their strength for causal inference.

Source: Fisher 1935

The limitations of randomisation include that it balances confounders only on average, so with small samples chance imbalances can still occur, which stratified or block randomisation can reduce; that it ensures comparability at baseline but must be protected by allocation concealment and maintained by proper conduct; and that it does not address biases arising after randomisation, such as differential dropout, which require other measures. Randomisation also does not guarantee generalisability. These limitations mean randomisation is combined with adequate sample size, concealment, blinding, and appropriate analysis to realise its benefits fully.

Source: Friedman, Furberg & DeMets 2015