Correlation and causation
On this page
Definition
Correlation is a statistical association: two quantities move together. Causation claims that changing one makes the other change. Observational data gives the first; only an intervention with a control gives the second.
How to compute
Test an observed association for confounders: a common factor driving both quantities, or reverse direction. Remove the bias with randomisation (an A/B test), instrumental variables or control methods, and read the effect size and its uncertainty from an interval.
Pitfalls
The textbook case is seasonality and selection: ice-cream sales and drownings rise together, but the shared cause is summer. Regression controlling for observed variables does not rescue you from unmeasured confounders, nor does it yield a causal estimate. Regression to the mean after a spike or a dip also looks like an intervention effect.