Causal Inference

From a statistical perspective, causal inferences are strongest when drawn from a “randomized controlled experiment,” where the investigators assign subjects at random to a treatment condition or a control condition. Up to random error, randomization balances the two groups with respect to all factors—except for the particular causal factor under investigation. A significant difference between the two groups is therefore evidence of causation. For instance, one can assign heart attack patients to take aspirin (treatment) or not (control). Patients in the treatment group are less likely to experience adverse events. The difference shows that aspirin is protective.

Managing experiments is often complicated and expensive; sometimes, experiments are impractical or unethical. Investigators may turn instead to observational studies. In such a study, subjects assign themselves to [Page ...