A cohort study identifies groups that differ in some exposure and follows them, recording outcomes as they occur.
A cohort study identifies groups that differ in some exposure and follows them, recording outcomes as they occur. Because the exposure is recorded before the outcome, the sequence is established, which is more than most observational designs manage.
People ended up in their groups by choice or circumstance, so the groups differ in ways beyond the exposure. Statistical adjustment handles measured differences and can do nothing about unmeasured ones, which is the design's central limitation.
Large cohorts followed for years can detect rare outcomes and long-delayed effects that no trial would ever be funded to find. For many questions this is the only evidence that will ever exist.
People who leave a study differ from those who stay, often in ways connected to the outcome. High attrition undermines a cohort more quietly than it undermines a trial, because there is no randomisation to appeal to.
Many cohorts record an exposure at enrolment and treat it as fixed for decades. Behaviour changes, and a result attributed to the baseline exposure may belong to something that happened later.
A prospective cohort decides what to measure before outcomes occur. A retrospective one assembles records after the fact and inherits whatever those records happened to contain. The word cohort covers both and they are not equally strong.
Several major causal conclusions in medicine rest on cohort evidence, because randomising the exposure would have been unethical. Consistency across many independent cohorts, with a dose-response pattern, is the shape that carried them.
How many were lost, whether exposure was re-measured, which confounders were adjusted for, and whether the authors name the ones they could not measure.
Where a period during which an outcome could not have occurred is misassigned to one group, that group gains an artificial advantage. It arises easily in record-based cohorts when exposure is defined by something that takes time to happen, and it has produced several large apparent benefits that disappeared on reanalysis.
People who adopt one health-related behaviour tend to adopt others and to differ in ways no dataset records. Cohorts comparing adopters with non-adopters inherit all of it, which is why observational findings about voluntary behaviours have a poor record of surviving randomised testing.
The claims on this page about following groups forward, and where the design is strong and weak can be checked against sources outside this site: