Why does correlation not imply causation?

Two variables can move together for reasons other than one causing the other: a third variable (a confounder) may drive both, the causation may run the opposite way (reverse causation), or the pattern may be coincidence. A correlation alone cannot distinguish these, so it cannot establish cause.

What does it take to establish causation?

The strongest evidence comes from a randomised controlled experiment, where random assignment breaks the link between the treatment and confounders. Where experiments are impossible, researchers use longitudinal designs, statistical control for confounders, and frameworks like the Bradford Hill criteria to build a careful causal case, but the bar is much higher than a correlation.

Can a correlational study ever suggest causation?

It can generate and support causal hypotheses, especially when the association is strong, consistent, dose-dependent, and temporally ordered, and when plausible confounders are controlled. But on its own it cannot prove causation, and your write-up should describe an association, not a cause.

Home › Guides › Statistics › Correlation vs causation

Correlation vs causation

Q: What is a confounding variable?

A confounder is a third variable associated with both the supposed cause and the supposed effect, creating a correlation between them without a direct causal link. The classic example is ice-cream sales and drownings, both driven by hot weather. Confounders are the main reason correlations mislead.

By Dr. Rafiq Muhammad, MD, PhD · Updated June 2026

“Correlation is not causation” is the most-quoted line in statistics — and the most-ignored. A correlation means two things move together; causation means one makes the other change. The gap between them is where bad conclusions live, because a correlation is consistent with several explanations and can’t tell you which one is true.

Why a correlation can’t prove cause

If A and B are correlated, at least four things could be going on:

A causes B — what you hoped.
B causes A — reverse causation (you have the arrow backwards).
A third variable C causes both — a confounder.
Coincidence — pure chance, especially with small samples or many comparisons.

A correlation coefficient can’t distinguish these. That’s the whole problem.

The classic example: ice-cream sales correlate with drowning deaths. Ice cream doesn’t cause drowning — hot weather drives both. That hidden driver is a confounder, and confounders are why most misleading correlations mislead.

The two traps to name out loud

Confounding — a third variable explains the link. Always ask: what else could be driving both?
Reverse causation — does happiness cause exercise, or exercise cause happiness? Without a time order, you can’t say.

What it actually takes to claim causation

The gold standard is a randomised controlled experiment: random assignment severs the link between the treatment and any confounders, so a difference in outcomes can be pinned on the treatment. When you can’t randomise, you build a causal case more carefully — longitudinal timing, statistically controlling for confounders, and weight-of-evidence frameworks (e.g. the Bradford Hill considerations). The bar is far higher than “they correlate.”

Write it honestly

If your study is correlational, your discussion should report an association, name the plausible confounders, and resist the slide into causal language (“X leads to Y”). Reviewers watch for exactly this overreach — and writing it precisely is a mark of a careful researcher.

Deciding whether your design can even support a causal claim? Start with types of research design — experimental vs correlational — and the Research Question Validator to pressure-test what you’re entitled to conclude.

Get the free Statistics toolkit

A reporting checklist, worked examples, and test-selection flowcharts from the Statistical Test Selection Workbook. We’ll email you the download link.

Get the Statistical Test Selection Workbook

Frequently asked questions

Why doesn't correlation imply causation?

Because A↔B could be A→B, B→A, a shared cause (confounder), or coincidence — a correlation can’t tell which.

What is a confounding variable?

A third variable that drives both, creating a correlation with no direct link — like hot weather behind ice cream and drownings.

What does it take to claim causation?

Ideally a randomised experiment; otherwise timing, controlling for confounders, and weight-of-evidence frameworks.

Can a correlational study suggest cause?

It can support a causal hypothesis, but on its own report an association, not a cause.

Types of research design → Open the Test Selector →