What Is a Randomised Controlled Trial? Letting Chance Do the Hard Work

Suppose a new treatment is given to patients whose doctors judge them suitable. Those patients will differ systematically from the ones who did not receive it, in severity of illness, in age, in other conditions and in a hundred unmeasured ways, and any difference in outcome afterwards could be caused by any of those rather than by the treatment. That is confounding, and it is the reason observational comparisons mislead so persistently. Random assignment breaks the link between who receives the treatment and every other characteristic at once, known and unknown, so that on average the two groups differ only in what was assigned. It does not guarantee balance in any single trial, since chance can produce an imbalance, and it guarantees that any imbalance is random rather than systematic, which is exactly what statistical tests are built to account for. That is why randomisation is not merely a fair way of allocating scarce treatment but the thing that makes the resulting numbers interpretable.

Randomisation alone is not enough, and the accompanying features each block a specific route by which a false result can arise:

• •Allocation concealment, so that whoever is enrolling participants cannot know or influence which group the next person will receive, which is different from blinding and is the safeguard most often done badly
• •A control group receiving either a placebo or the current standard treatment, since improvement over time happens anyway through natural recovery and regression to the mean
• •Blinding of participants, which prevents expectation affecting reported outcomes, and of clinicians and assessors, which prevents it affecting treatment and measurement
• •A prespecified primary outcome, since allowing the outcome to be chosen afterwards permits picking whichever measure happened to move
• •Intention to treat analysis, counting everyone in the group they were assigned to regardless of whether they complied, because excluding dropouts reintroduces exactly the selection randomisation removed
• •Adequate sample size calculated in advance, since an underpowered trial cannot distinguish a real effect from nothing
• •Trial registration before enrolment, which makes selective reporting and unpublished negative trials detectable

The method has real limits and treating it as the only acceptable evidence causes its own problems. Some questions cannot ethically be randomised, since nobody will assign people to smoke or to be malnourished, which is why the link between smoking and lung cancer was established from observational evidence carefully analysed. Rare outcomes and long-term harms require sample sizes and durations beyond what trials can manage, so post-marketing surveillance does that work. Trial populations are frequently unrepresentative, excluding the elderly, the pregnant and those with multiple conditions, which limits how far results transfer to the patients actually treated. Short trials measure surrogate outcomes, such as a laboratory value, which may not track the outcome that matters. Complex interventions like surgery, education programmes or policy are hard to blind and hard to standardise. And a trial answers whether something worked on average in that population, which is not the same as whether it will work for a particular person.

The method spread well outside clinical research and the spread has been consequential. Development economics adopted field experiments to test interventions including microcredit, deworming, teacher incentives and cash transfers, work recognised with a Nobel prize in 2019 and criticised on the grounds that a result obtained in one district at one time may not transfer, that randomising social programmes raises ethical questions, and that the method suits small questions better than structural ones. Technology companies run continuous randomised experiments on their users, usually called split testing, at a scale that makes them the largest experimental enterprise in history and with almost no external oversight. Governments have established units running trials on policy interventions from tax letters to job centre procedures. The common pattern is that randomisation is powerful wherever an intervention can be assigned, and that the hardest questions in every field tend to be the ones where it cannot.

Random assignment breaks the link between receiving a treatment and every other characteristic at once, including unknown ones, which is what makes the comparison interpretable and is why deliberate matching cannot substitute. Allocation concealment, blinding, a prespecified outcome, intention to treat analysis and prior registration each block a particular route to a false result. Trials cannot answer questions that cannot be ethically randomised, and their populations are frequently unrepresentative of patients actually treated.