Correlation vs. Causation: How to Tell the Difference

A correlation tells you that two things move together; causation tells you that changing one would change the other. That is the whole of correlation vs causation: when A is high, B tends to be high (or low), but only causation predicts what happens to B if you act on A.¹

A pattern has at least five possible explanations, and “A causes B” is only one. B may cause A, a third factor may drive both, the match may be luck, or the way people got into the data may have produced it. The questions for sorting them out mostly go back to a 1965 lecture by the British medical statistician Austin Bradford Hill. Hill’s own verdict was that none of his nine viewpoints can prove cause and effect on its own, and none is required; together they ask whether any other explanation is as likely.²

When a headline, a report or a dashboard says one thing drives another, five quick questions do most of the work:

1. Did anyone assign the change at random, or did people choose it?
2. Which came first?
3. What third factor could sit behind both?
4. Was the link found by trawling through many numbers?
5. Who is missing from the data?

Prefer to watch? The video covers the same ground as this article.

VideoCorrelation vs Causation: How to Tell the DifferenceWiserHours

Correlation vs causation: what each word claims

A correlation measures how closely two things move together; causation means one of them has an effect on the other. The Association of Health Care Journalists (AHCJ), a US professional body, puts it this way in its glossary for reporters: two variables can be related without one being caused by the other, and studies that only observe people can show correlation, not causation.¹

A correlationcorrelation: A measure of how closely two things move together, running from minus one, where one rises as the other falls, through zero, meaning no link, to plus one. It says how strong the relationship is, not what causes it, and it is not a percentage.Full entry in the glossary is usually reported as a number between minus one and plus one. It tells you how tightly the two measures track each other, and nothing about why. A value near plus one for umbrella sales and late trains looks the same on paper as a value near plus one for a training course and later promotions, even though the stories behind them could not be more different.

The words around a finding carry the claim. The AHCJ glossary asks reporters to keep the word “effect” for cases where causation is involved, because readers take “effect” to mean the exposure was the cause.³ Its entry on correlation suggests that a story on an observational study say people who sleep less are “more likely” to gain weight, not that short sleep “may cause” weight gain.¹

When a headline, a report or a colleague uses a causal verb, check whether anyone changed anything on purpose. If the evidence is a survey, a spreadsheet or a comparison of people who made different choices, read the verb as “goes with” until you know more. It is the same pause that guards any fast first impression, a theme of why our judgment takes shortcuts and of making sounder decisions in general.

Why correlation does not imply causation: four other explanations

Correlation does not imply causation because the same pattern can come from four other sources: reverse causation, confounding, chance and selection. Each has a question that tests for it. Treat the four as a portable checklist, the kind of thinking tool you can carry from one problem to the next.

1. Reverse causation: B drives A: the "effect" turns out to be the cause
2. Confounding: a third factor drives both, and A and B only look connected
3. Chance: search enough data and some pairs line up by luck
4. Selection: who ends up in the data creates a link that the wider group does not have

Reverse causation: the “effect” may be the cause

Reverse causation means the arrow points the other way. The AHCJ’s example is diet soda: some studies found that people who drink it are more likely to be obese, which led to speculation that artificial sweeteners somehow cause obesity. Critics answered that people who are gaining weight may switch to diet drinks to cut calories.⁴

Hill called this the question of temporality, “which is the cart and which the horse?” His own example came from industry: does a job raise the risk of tuberculosis, or do people who are already more liable to it select that kind of work?² Picture a manager who notices that projects with more status meetings miss more deadlines. The meetings may slow things down, or troubled projects may simply attract more meetings.

What to check: which came first, and could the outcome have prompted the supposed cause? Data collected at one point in time cannot answer that; data that follow the same people or teams over time can at least show the order.

Confounding: a third factor behind both

A confounding variable is a third factor that affects both the supposed cause and the outcome, so the two look linked when neither drives the other. The AHCJ glossary gives a clean example: later-born children were found to have a much higher risk of Down syndrome than first-born children, but much of that link was explained by the mother’s age, since mothers having a third or fourth child also tend to be older.⁵

Confounders hide in workplace data too. Suppose staff who use the company gym take fewer sick days. People who are already in good health, or whose jobs leave room for a lunchtime workout, are more likely both to use the gym and to stay well, so the gym may get credit that belongs to them. Researchers try to adjust for confounders they can measure, but the Cochrane Handbook, the methods manual for Cochrane’s medical evidence reviews, warns that in studies without random assignment, unmeasured confounding can usually not be excluded, because researchers are seldom sure they know every confounder.⁶

What to check: name the most likely third factor out loud, then ask whether the comparison held it steady. If nobody can name one, look harder; age, income, health, seniority and the team’s manager are common culprits.

Chance: some pairs line up by luck

Some correlations are coincidences that look meaningful because someone searched until they found one. Tyler Vigen’s Spurious Correlations project, a deliberately playful illustration built from public US data, shows per-person butter consumption in the US rising and falling with the number of epidemiologists in Colorado from 2005 to 2020, with a correlation of 0.92.⁷

Vigen explains the trick: his database holds more than 25,000 variables, and he compares every one against every other. This is data dredging, and any large enough dataset will throw up strong correlations at random. He also warns that a confounding event such as a global pandemic can make unrelated series look connected.⁷

Workplace dashboards invite the same mistake. Track dozens of metrics across many teams and some pairs will line up every quarter by luck alone. Hill’s advice was that whether a link is chance or real can sometimes be settled only by repeating the observation.² Be most skeptical of a striking link that was found by browsing, and see whether it survives in next quarter’s data before you act on it.

Selection: who ends up in the data

Selection bias arises when the way people enter a dataset creates a link that the wider group does not have. Hill passed on an example from a 1958 paper by J. A. Heady: patients admitted for a peptic ulcer were asked about recent crises in their lives and compared with patients admitted for a simple hernia. The groups were not alike, he noted: a perforated ulcer sends you to hospital at once, crisis or not, while someone with a hernia might stay home for a while after a personal upheaval. Hill saw the same problem with volunteers, since only an unknown fraction of factory staff volunteer for a procedure or interview.²

An office might find that employees who took an optional leadership course were promoted more often. People who sign up for optional courses may be more ambitious or have more supportive managers. That is self-selectionself-selection: When the people in a study chose to take part, or chose the thing being studied, rather than being picked or assigned. Volunteers and choosers can differ from everyone else, so the result may not apply more widely or may reflect who chose rather than what they chose.Full entry in the glossary at work: the course may be getting credit for who chose it.

Myth

A big enough dataset settles whether a link is causal.

Fact

More data shrinks the role of chance, not of selection or confounding. Hill wrote that no number of exact repetitions would remove a selection fallacy.

What to check: ask who is missing. People who did not answer the survey, customers who left, projects that were cancelled and staff who quit before the review all vanish from the data, and they rarely vanish at random.

A workplace example: engagement scores and performance

Employee engagement and performance are a real link with an unsettled direction. Gallup’s 2024 Q12 meta-analysis, a company report built on studies of Gallup’s own survey, finds that business units with higher engagement tend to perform better, but states that the paper does not directly address causality. Gallup adds that its earlier causal studies point to a two-way relationship, in which engagement predicts financial results more strongly than the reverse.⁸

An independent review reads the evidence more cautiously.

The study

Moderate evidence

Six reviews, 23 single studies and one randomized trial

A 2021 rapid evidence review by the CIPD, the UK-based professional body for HR and people development, appraised six meta-analyses or systematic reviews and 23 longitudinal single studies. Engaged staff did tend to perform better, mostly measured person by person, yet the effects were consistently small, and several studies found the arrow running the other way as well: doing well can lift engagement. Just one of the 23 single studies assigned people at random.⁹

Now run the four explanations on a dashboard where high-engagement teams also hit their targets. Hitting targets may lift morale (reverse causation). A good manager may raise both (confounding). With dozens of teams and metrics, some pairs will match by luck (chance). And if struggling teams answer the survey less often, the link is partly about who replied (selection). A manager learns more by treating an engagement program as a test: try it with some teams, keep comparable teams as a comparison, assign them at random if you can, and compare results over several quarters.

How do you prove causation?

Causation is never proved beyond all doubt; the best evidence for it comes from a randomized controlled trialrandomized controlled trial: A study that assigns participants to the treatment or the comparison group at random, so the groups start out alike and a difference in what happens next can be put down to the treatment. Random assignment makes the two groups comparable; it does not make the people in the trial representative of anyone else.Full entry in the glossary, where chance alone decides who gets the change. The Cochrane Handbook explains why: randomization stops known and unknown factors that predict the outcome from influencing who gets the change, so those factors cannot become common causes of both.⁶ Because the change is made before the outcome is measured, the outcome cannot have caused it.

Many questions cannot be randomized, from smoking to company culture. For those, Hill offered his nine viewpoints in 1965, and stressed that they are aids to judgment, not rules.

None of my nine viewpoints can bring indisputable evidence for or against the cause-and-effect hypothesis and none can be required as a sine qua non.

Two of his examples show how the questions work. On strength, he recalled John Snow’s analysis of the 1854 cholera epidemic: houses supplied by a grossly polluted water company had a death rate 14 times that of houses supplied by a rival company with sewage-free water. On dose-responsedose-response: A pattern in which more of an exposure goes with more of an outcome, step by step. In observational research it is one of the signs that makes a causal link more plausible, though on its own it does not prove cause.Full entry in the glossary, he noted that lung cancer deaths rose steadily with the number of cigarettes smoked each day.²

For everyday use, the questions that carry most of the weight fit on a card.

A card for the next 'causes' claim you meet

• Was the change assigned at random, or did people choose it?
• Did the supposed cause come before the effect?
• Is there a dose-response: does more X go with more Y?
• Has the link been found again, by others and with other methods?
• Is there a plausible mechanism, and does it fit what else is known?
• Could a third factor, luck or who is in the data explain it instead?

How to read a “linked to” claim in a headline or a report

Reading a “linked to” claim well takes four quick steps: translate the verb, find the comparison, run the four alternatives, and size your response to the cost of being wrong. None of it needs statistics training.

1. 1Translate the verblinked to means goes with
2. 2Find the comparisonwho was compared, and how they got there
3. 3Run the four alternativesreverse, confounding, chance, selection
4. 4Size your responseto the cost of being wrong

“Linked to”, “associated with” and “more likely” describe a pattern; “boosts”, “drives” and “causes” claim more, and are fair only once causation has been shown, which usually takes an experiment.³ If people chose which group they were in, selection and confounding are live explanations.

The last step is the one people skip. Hill argued that the evidence needed before acting depends on what acting would cost: fairly slight evidence can justify a change that does little harm if it proves wrong, while much stronger evidence is needed before asking people to give up something they value.² A team that tries a no-meeting morning because it is linked to better focus risks little. A company that restructures pay because of a correlation risks a great deal.

When the answer is “very little”, a small test is often the fastest way to learn. When the answer is “a lot”, wait for evidence that has been through Hill’s questions, ideally a randomized comparison.

The bottom line

Two measures rising together are a clue to look into, and only evidence about what happens when one of them changes tells you about cause. Before you accept “X causes Y”, check whether the arrow could run backwards, whether a third factor or chance could explain it, and who is missing from the data. Where you can, test the change with a randomly chosen comparison group, and ask for more evidence the more it would cost to be wrong.