Research
Observational evidence
What observational studies contribute that randomised trials cannot, and where their conclusions can mislead.
- Last updated:
- 2026-07-22
Overview
Observational studies (cohort studies, case–control studies, registries, real-world evidence) analyse outcomes in populations without randomising the intervention. They can capture patterns at a scale and duration that randomised trials cannot practically match — millions of person-years, decades of follow-up, uncommon adverse events.
They are also susceptible to confounding: differences between people who did and did not receive the intervention that themselves cause the observed outcome. Statistical adjustment reduces confounding but does not eliminate it.
Why it matters
Some of the most important safety information about approved medicines comes from post-marketing observational data. Some of the most misleading efficacy claims come from the same source.
Key concepts
Association vs. causation
Observational data reliably establish associations. Causation requires either randomisation or very careful additional reasoning (e.g. Mendelian randomisation, natural experiments).
Confounding
The classic pitfall: people prescribed a medicine often differ systematically from people not prescribed it, and those differences may drive the outcome.
Rare-event detection
Observational data are often the only feasible way to detect rare adverse events; trials are usually too small.
Common misconceptions
- "Big data proves it." Statistical significance in a huge dataset does not resolve confounding; it can amplify it.
- "Real-world evidence beats trial evidence." It provides different evidence; it does not replace trial evidence for questions trials can answer.
- "A registry that shows no adverse events is proof of safety." A registry may simply not be capturing them.
Practical interpretation
When a compound page cites observational data, note whether it is for benefit or for harm. Observational evidence is generally more trustworthy for the harm question than for the benefit question.
Limitations
Study quality varies. Peer review, pre-registration, and independent replication improve trust; their absence should raise it.