Correlation is not cause: a practical check

Correlation means two measures tend to move together: when one is higher, the other is usually higher too, or usually lower. That pattern is useful, but it does not show that one causes the other. Establishing cause requires more evidence, because the link may run through a third factor, may run in the opposite direction, or may be a coincidence within the data you have.

A classic example makes the problem clear. Ice cream sales and the number of drowning incidents both rise in the warm months and fall in the cold ones. Neither causes the other. Hot weather drives both, and that shared driver is a common pattern in real data. Anyone who sees the two lines climb together and concludes that ice cream is dangerous has mistaken a shared cause for a direct one.

Start with what the numbers actually show

Before asking about cause, describe the relationship precisely. How strong is it? A scatter plot or a simple table of paired values shows whether the two measures move together closely or only loosely. A weak relationship with a lot of scatter supports a much smaller claim than a tight one. Look at the spread around any apparent trend as well. A tight cluster and a loose cloud can share the same overall direction, yet they justify very different levels of confidence. Check the direction, too, and whether the pattern holds across the whole range of values or only at the extremes.

Check how each measure was defined and collected. If one number is a count and the other is a rate, or if they cover different areas, age groups or time spans, the relationship may reflect how the data were built rather than anything happening in the world. Writing down exactly what each number measures keeps the later steps honest, because it is easy to forget the definitions once a chart looks convincing.

Check the timing

A cause has to come before its effect. If a change in one measure is consistently followed by a change in the other, that is a useful clue. If the two move in the same period, timing alone cannot tell you which came first; if the supposed cause shows up only after the effect, the timing argues against it.

Timing is a clue, not proof. Many things that happen in sequence are unrelated, and some causes act with a long delay. Look at the pattern over several cycles rather than a single event, and ask whether the lag between the two measures makes sense for the mechanism you have in mind. A lag of minutes is plausible for some effects and implausible for others.

Look for a third factor

The most common explanation for a misleading correlation is a confounder, a third factor that influences both measures. Temperature is the confounder in the ice cream example. In everyday data, income, age, location and the season often play this role. A short list of candidate factors, each tested with the question "could this affect both measures?", is a practical way to surface them.

Then test the candidate. Compare groups that are similar on that factor. If the relationship between your two measures mostly disappears when you look only at days with similar temperatures, or only at people in a similar age band, the confounder was doing much of the work. If the relationship survives the comparison, it is more believable, though still not proven.

Consider reverse causation and chance

Sometimes the arrow points the other way. People who feel unwell may buy more medicine, but the medicine did not make them unwell. Students who study more may score higher, and higher scores may also encourage more study. Ask which measure could plausibly be the cause, and whether the proposed direction would produce the same pattern if it were reversed.

Chance deserves a check, particularly when the data set is small. A handful of paired values can show a strong-looking pattern by accident. Asking whether the same pattern appears in a fresh set of observations is a reasonable test. If you have only one study or one season of data, treat a strong causal claim with caution until someone repeats the finding under different conditions.

Run through a short checklist

When someone says one thing causes another, a few questions give a quick and orderly assessment:

• Does the relationship hold when it is measured again, in a different group or period?
• Does the proposed cause come before the effect, after a plausible interval?
• Could a third factor explain both measures?
• Could the direction be reversed?
• Is there a mechanism that accounts for the link, rather than only a pattern?

Write the answers down rather than running through them in your head. Writing forces you to say what the evidence actually is, which makes gaps far easier to see. A yes to all five does not prove causation, but it makes the claim more plausible. A no to any of them is a reason to hold back. Randomized controlled experiments, in which chance decides who receives the factor, are the strongest way to establish cause, but they are often impractical or unethical for everyday questions. When a claim matters to a decision, decide in advance what evidence would change your mind, and check whether you have it.