Randomly assigning participants tends to distribute every characteristic evenly between groups, including ones nobody measured or thought of.
Randomly assigning participants tends to distribute every characteristic evenly between groups, including ones nobody measured or thought of. No amount of statistical adjustment afterwards achieves this, because adjustment can only handle variables that were recorded. That is the entire argument for the design.
If people choose their own group, the choice itself carries information. Healthier, wealthier or more motivated people select differently, and the resulting difference between groups is then attributed to the treatment rather than to the people who chose it.
Randomisation can be undone if whoever enrols participants can see the next assignment and delays a sicker patient to the other arm. Proper trials conceal the sequence from the person enrolling, and a paper should state how that was done rather than simply using the word randomised.
The word appears in the title of many weak trials. Randomisation addresses selection and does nothing about short follow-up, surrogate endpoints, or selective reporting of outcomes. It is a necessary condition for a strong trial rather than a sufficient one.
Sometimes whole clinics or regions are randomised instead of individuals. People within a cluster resemble each other, so the effective sample size is smaller than the participant count suggests. Analyses ignoring clustering report intervals that are too narrow and significance that is too easy to reach.
With enough measured characteristics, some will differ between groups by chance. One or two imbalances across twenty comparisons is normal behaviour. Testing each for significance is a known error, because randomisation guarantees the process rather than the outcome of any single draw.
Some questions cannot be randomised for practical or ethical reasons. Observational work is then the only evidence available, and reading it carefully beats dismissing it — but the ceiling on what it can establish is lower, and an honest paper says where that ceiling sits.
How the sequence was generated, who held it, whether the people assessing outcomes knew the assignment, and whether the baseline table shows the groups actually came out comparable.
Where a few factors are known to matter a great deal, trials sometimes randomise within strata or use minimisation to balance as they enrol. Both preserve the essential property while improving balance on the named factors. Analyses should account for the method used, and papers that stratify and then analyse as if they had not are reporting intervals that are slightly too wide.
A trial that randomised in blocks, by cluster or with stratification has a structure the analysis must respect. Using a simpler model than the design warrants is a common and quiet error, and it is visible only if the methods describe both the allocation and the analysis clearly enough to compare.
The claims on this page about what random allocation controls for, including the things nobody thought to measure can be checked against sources outside this site: