observational study

The Centre for Evidence-Based Medicine at the University of Oxford draws the line by what the researchers do. Observational studies investigate and record exposures, such as interventions or risk factors, and observe outcomes such as disease as they occur; in an analytic observational study the researcher simply measures the exposure or treatments of the groups. In an experimental study the researcher manipulates the exposure instead, allocating who receives what. Cohort studies, case-control studies and cross-sectional studies are all observational. The centre says of the cross-sectional kind that it establishes association at most, not causality, and of the cohort kind that the exposure may be linked to a hidden confounder.

That hidden factor is the standing problem, and it follows directly from the missing allocation. The Cochrane Handbook explains that randomisation aims to balance prognostic factors across intervention groups, thus preventing confounding, which occurs when there are common causes of intervention group assignment and outcome. Take the randomisation away and nothing guarantees the groups were alike to begin with. The handbook puts the consequence plainly: potential biases are likely to be greater than in randomised trials, because some of the protections against bias available to a trial are not established here, and few circumstances are likely to give a similar level of protection against confounding as randomisation does.

None of that makes the design weak or optional. It is the only design available for most questions about how people actually live, where allocating an exposure would be unethical or impossible, and it can follow far larger groups for far longer than a trial can afford. What it cannot do is tell you which way the arrow runs. The Cochrane Handbook prefers the label non-randomised study of interventions and discourages the word observational for it, on the grounds that the term is used in various ways, so a reader meeting either phrase should read it as the same warning: the researchers watched, they did not assign.