Selection bias – some ‘equal groups’ are more equal than others
The groups of participants receiving interventions should be equal, otherwise confounding variables might give one treatment an advantage over another. If there is a systematic reason for this, the study is at risk of selection bias.
Randomization (sequence generation)
The first consideration is whether the treatments were allocated randomly. The best method is to predefine a randomization schedule, in which treatments are allocated by chance alone. Other methods could introduce differences between groups. An example with obvious implications (to highlight the point) would be to randomize preterm boys to one group, and girls to another. It might seem reasonable to allocate treatments, after participants enrol, by flipping coins or rolling dice, but this might make group sizes unequal. ‘Quasi-randomisation’ usually implies that patients receive treatments based on pre-randomisation assessment, which by definition introduces differences between groups, leading to particularly high risk of bias.
Allocation concealment
The second consideration is whether the person recruiting to the trial knows which treatment the next participant would receive, as they might (consciously or subconsciously) influence group composition. For example, if I know that the next patient does not take treatment reliably, I may be reluctant to randomize them to the exciting experimental treatment. Alternatively, for clinical reasons, I might want them to receive it.
Keeping the identity of the next randomized treatment hidden from the recruiter is called allocation concealment, and is crucial. Sir Iain Chalmers describes the MRC’s controlled trial of streptomycin for tuberculosis as landmark not because it was randomized, but “…because of the clearly described precautions that were taken to conceal the allocation schedule from those involved in entering patients”.
Adequate methods of concealment preclude the recruiter from identifying the next allocation, and include centrally held randomization schedules, sequentially numbered drug containers, or sealed opaque envelopes (which cannot be held up to the light).
(Another one by @Iansinah)