Case-Control Study

A case-control study works backwards from outcome to exposure. Investigators assemble people who have the condition of interest, a comparison group who do not, and reconstruct past exposure in both. Because sampling is conditioned on the outcome, the design cannot estimate incidence or absolute risk; its natural measure is the odds ratio, which approximates the risk ratio only when the outcome is rare in the source population. A nested case-control study applies the same logic inside an existing cohort.

The design earned its place with rare outcomes and long latencies. The 1950 work of Doll and Hill linking smoking to lung cancer is the canonical example, and it remains the workhorse for drug safety questions where events are uncommon: a French nationwide case-control analysis published in 2023, using national claims data, reported an association between longer cumulative use of GLP-1 receptor agonists and thyroid cancer.

The reason to run one is efficiency. To study an outcome occurring in a small fraction of users, a cohort would follow an enormous number of people for years, whereas a case-control study needs only the cases already accrued plus a manageable set of controls. What it buys in feasibility it pays for in fragility, since everything depends on whether controls came from the same population that produced the cases.

Two failure modes dominate. Recall bias arises when people with a diagnosis search their memory harder for explanations than healthy controls do, which is why dispensing records beat interviews wherever they exist. Protopathic bias arises when early symptoms of the undiagnosed condition prompt the exposure, making a drug look causal when it was prescribed in response to the disease.