Correlation vs. Causation: How to Tell the Difference

Ice-cream sales rise when sunburns rise. One does not cause the other; hot, sunny weather helps produce both. This familiar example is simple because the third factor is easy to see. Real research is harder. Confounders may be partly hidden, causes may run in both directions and the data may come from people who differ in important ways before a study begins.

What correlation tells us

A correlation describes a pattern. When one quantity is higher, another tends to be higher, lower or otherwise predictably different. The pattern may be strong or weak, linear or curved, stable or present only in a subgroup. Correlation can be scientifically valuable: it reveals regularities, generates hypotheses and enables prediction even before a mechanism is understood.

But prediction and intervention are separate questions. If a marker predicts an outcome, using the marker to identify risk may help. Changing the marker will improve the outcome only if the marker lies on the relevant causal path—or the intervention changes something that does.

Five explanations for the same pattern

Confounding: the third-variable problem

A confounder is associated with both the proposed cause and outcome and can distort their relationship. Suppose a study finds that people who use a particular app have different learning outcomes. Age, previous knowledge, income, motivation, school resources or time available could influence both app use and performance. Statistical adjustment can help only for confounders that were measured adequately and modeled sensibly. It cannot guarantee that every important difference disappeared.

Researchers therefore begin with causal knowledge, not a machine that indiscriminately “controls for everything.” Adjusting for a variable caused by the exposure can remove part of the effect being studied. Adjusting for a common consequence can even create an association. The list of variables in a model is not a quality score.

Why time order matters

A cause must precede its effect, but datasets often measure everything at once. If a survey finds that poor sleep and worry occur together, it cannot by itself determine direction. Worry may disturb sleep; insufficient sleep may intensify worry; both may reinforce each other; or another condition may affect both.

Longitudinal studies follow observations through time and can establish sequence more clearly. They still face confounding, dropout and changes in measurement. Time order is necessary for causation, not sufficient.

Randomized experiments and causal inference

Random assignment gives each participant or unit a defined chance of receiving an intervention. When conducted properly, it tends to balance both measured and unmeasured background factors between groups. A difference that appears after assignment can therefore be attributed more confidently to the intervention.

Randomization is not magic. Non-adherence, missing outcomes, unblinded measurement, small samples or an unsuitable comparison can weaken an experiment. Results may apply only to the tested population, dose, setting and period. Some questions cannot be randomized because doing so would be unethical or impossible.

How observational evidence can support causation

Much of astronomy, climate science, geology, ecology and population research depends on observation. Strong causal reasoning can emerge when several features align: the cause precedes the effect; larger exposure predicts larger response; alternative explanations fail; natural experiments create informative comparisons; a plausible mechanism exists; and different methods produce compatible findings.

No checklist mechanically proves causality. The point is triangulation. Every method has weaknesses, but different methods often have different weaknesses. When their answers converge, it becomes harder for one bias to explain everything.

A worked example: umbrellas and rain

On rainy days, umbrellas and wet roads are correlated. Umbrellas do not make roads wet. Rain causes both. Now imagine a city tests whether free umbrellas reduce how soaked pedestrians become. If umbrellas are randomly offered at some transit exits and not others, and wetness is measured consistently, the experiment addresses a different causal question: the effect of access to umbrellas on pedestrians under rainy conditions.

The example shows why variables are not inherently “causal” or “noncausal.” The exact question matters. Rain causes umbrella use; access to an umbrella can cause less exposure to rain. A good claim names the intervention, outcome, population, comparison and time.

Warning signs in headlines

• The study is observational, but the headline uses causes, prevents or leads to.
• No plausible alternative explanations are discussed.
• The sample is self-selected or very different from the population named in the headline.
• The relationship appears only after many outcomes or subgroups were searched.
• A relative change is reported without the underlying absolute numbers.
• A biological mechanism is treated as proof rather than one piece of evidence.

Seven questions for any causal claim

1. What exactly are A and B? Replace vague labels with measured variables.
2. Which came first? Establish timing.
3. Compared with what? Look for a credible counterfactual: what would have happened otherwise?
4. What could cause both? Identify major confounders.
5. How were cases selected? Selection can manufacture patterns.
6. Would another method agree? Seek experiments, natural experiments, longitudinal evidence or mechanisms.
7. How large and uncertain is the effect? Causal does not automatically mean important.

Connect the idea to the wider evidence

Correlation is not “bad evidence.” It is evidence whose meaning depends on design. Read What Does Statistically Significant Really Mean? to separate detectability from importance, and How to Read a Scientific Study to inspect methods and figures. Our cornerstone guides—How Scientific Discovery Works and How to Evaluate New Scientific Discoveries—place causal claims inside the full research process.