Correlation vs Causation

The Fundamental Distinction

Correlation means two variables move together — when one increases, the other tends to increase (or decrease). Causation means one variable directly produces a change in another. Correlation is necessary for causation but is not sufficient to prove it.

Ice cream sales and drowning deaths are strongly correlated. Ice cream does not cause drowning — both increase in summer due to a confounding variable (hot weather). This example is obvious, but in public policy, confounding variables are often hidden and the stakes are much higher.

Confounders and Reverse Causation

A confounding variable is a third factor that influences both variables being studied, creating a false impression of a direct relationship. Countries with more Nobel Prize winners also consume more chocolate per capita — the confounder is national wealth, which funds both research institutions and chocolate imports.

Reverse causation is another trap. Studies show that hospital patients who receive more visitors recover faster. But visitors may not cause recovery — patients who are recovering may simply be more pleasant to visit. The causal arrow might point in the opposite direction from what the headline suggests.

Worked Example: A Confounder in Action

Across a year, ice-cream sales and drowning deaths in a coastal region move together almost perfectly. Suppose the monthly correlation coefficient is r = 0.92, close to a perfect 1.0. Does buying ice cream cause drowning?

The pattern is real, but the cause is a confounder: summer heat. Hot weather independently drives up both ice-cream sales and the number of people swimming, and more swimmers means more drownings. Ice cream and drowning are correlated only because both are effects of a third variable. Control for temperature — compare only days at the same temperature — and the ice-cream/drowning link vanishes.

This is why a correlation coefficient, however large, can never by itself establish causation. Three rival explanations must always be ruled out before you infer that X causes Y:

1. Confounding — a third variable Z causes both X and Y (the heat above).
2. Reverse causation — Y actually causes X (do hospitals make people sick, or do sick people go to hospitals?).
3. Coincidence — with enough variables, some will correlate by pure chance (spurious correlations).

The gold standard for cutting through all three is the randomised controlled trial, which randomly assigns who receives the treatment so that, on average, confounders are balanced between the groups. Where experiments are impossible, careful researchers use natural experiments, statistical controls, and the strength-consistency-plausibility criteria to build a causal case.

Bradford Hill in Brief

When experiments are impossible, epidemiologists lean on criteria that make a causal story more or less credible. Strength: bigger effects are harder to explain away by a hidden confounder. Consistency: the link recurs across different populations and study designs. Dose-response: more exposure yields more effect. Temporality: the cause must precede the effect — the one criterion that is non-negotiable. Plausibility: there is a believable mechanism. None of these alone proves causation, but together they let careful analysts distinguish a robust causal claim from a suggestive coincidence.