Explainer
Correlation vs causation: why linked doesn't mean caused
Correlation measures the size and direction of a relationship between variables, but it does not automatically mean that a change in one causes the change in the other.[1][2] A third factor can drive both, as when ice cream and sunscreen sales rise together because of the season, and in practice cause and effect is much harder to establish than correlation.[3][4] Controlled studies, in which comparable groups receive different treatments, are the most effective way to test for causation; when they are not possible, observational studies add evidence but carry a higher risk of confounding.[5][6][7]
“People who do X are more likely to have Y” is one of the most common kinds of statistic in the news. It describes a correlation. Whether X actually causes Y is a separate and harder question.[2][4] This page explains the difference using the definitions published by the Australian Bureau of Statistics (ABS) and the research methods guidance behind how epidemiology studies work.
What correlation measures
The ABS defines correlation as a statistical measure, expressed as a number, of the size and direction of a relationship between two or more variables.[1] For two variables it is measured by the correlation coefficient, written r, which runs from +1.0 to -1.0. A positive value means the variables tend to rise and fall together; a negative value means one tends to fall as the other rises.[8]
The coefficient has limits. It usually measures a linear, straight-line relationship, so a real relationship that is curved can produce a coefficient close to zero.[9] And the ABS says plainly that the coefficient should not be used to say anything about cause and effect.[10]
Correlation is not causation
Causation means one event is the result of the other.[11] A correlation does not automatically mean that a change in one variable causes the change in the other.[2] The ABS’s example: smoking causes an increased risk of lung cancer, and smoking is also correlated with alcoholism, but it does not cause alcoholism.[11] The ABS calls causality an area of statistics that is commonly misunderstood by people who assume a correlation means there must be a causal link.[10]
Ice cream sales and sunscreen sales rise and fall together through the year. Buying ice cream does not make people buy sunscreen; hot weather drives both.[3] Whenever two things move together, ask whether something else could be behind both.
Confounding: the hidden third factor
In the ice cream example, a common cause (the season) sits behind both trends. In research on treatments and outcomes, a common cause of this kind is called a confounder: the Cochrane Handbook, a standard guide to health research methods, says confounding occurs when there are common causes of both the choice of intervention and the outcome, so the association you observe differs from the true causal effect.[12] Observational studies, in which researchers measure outcomes without assigning a treatment, are especially exposed to it.[13] Under the GRADE system used by Cochrane, evidence from non-randomised studies generally starts two levels lower than randomised evidence because of the risk of confounding and selection bias.[7]
GRADE does allow certainty in non-randomised evidence to be rated up, for example when the effect is large, when there is a dose-response gradient, or when plausible confounding would only have weakened the observed effect.[14] A cross-sectional study, which collects data at one point in time, can suggest an association but does not prove a link.[15]
How researchers test for cause
The ABS says a controlled study is the most effective way to establish causality. The population is split into groups that are comparable in almost every way, the groups receive different treatments, and their outcomes are compared.[5] Evidence rankings such as the one used by the US National Cancer Institute put randomised controlled trials first, followed by non-randomised trials and then cohort or case-control studies.[16]
Controlled studies are not always possible. It would be unethical to make one group take up a harmful habit, so observational studies are often used instead, adding evidence to the wider process of deciding whether a causal link exists.[6]
When trials overturned observed links
Randomised trials have contradicted conclusions drawn from observational data. Hormone replacement therapy had been thought to prevent heart disease, but the Women’s Health Initiative, a randomised trial of more than 68,000 postmenopausal women run alongside an observational study of more than 93,000, found that it did not.[17] In two trials of beta-carotene supplements in heavy smokers, lung cancer incidence and deaths were higher in those given the supplement, and one trial was stopped early for harm.[18]
Questions to ask about a claimed link
- Is it a correlation or a tested cause? A correlation alone does not show that one thing causes the other.[2]
- Could something else drive both? Look for possible confounders, like the weather behind ice cream and sunscreen sales.[3][12]
- What kind of study was it? Randomised trials sit above observational studies in evidence rankings.[16]
- How big is the effect? Relative figures can make small changes look large; see relative vs absolute risk.[19]
- Is “significant” doing the work? A small p-value does not by itself show an effect is large or important; see p-values and statistical significance.[20]
Read next
Next in the crash course
P-values and statistical significance, explained
What a p-value is and is not, where the 0.05 threshold came from, common misreadings, and the American Statistical Association's six principles.
Questions readers ask
What is the difference between correlation and causation?
Correlation is a statistical measure of the size and direction of a relationship between variables. Causation means one event is the result of the other. A correlation does not automatically mean one variable causes the change in the other.[1][11][2]
What does a correlation coefficient tell you?
The correlation coefficient r runs from +1.0 to -1.0. A positive value means the variables move together and a negative value means they move in opposite directions. It usually measures only linear relationships, and it should not be used to say anything about cause and effect.[8][9][10]
What is a confounder?
Confounding happens when something causes both the choice of intervention and the outcome, so the association you see differs from the true causal effect. A simpler example of a common cause behind two trends, from the Australian Bureau of Statistics, is ice cream and sunscreen sales, which rise and fall together because of the season.[12][3]
How do researchers show that something causes something else?
The most effective way is a controlled study, in which a population is split into groups that are comparable in almost every way and that receive different treatments. Where that is unethical, observational studies add evidence to the wider process of judging whether a causal link exists.[5][6]
Has an observed link ever been overturned by a trial?
Yes. Hormone replacement therapy had been thought to prevent heart disease, but the Women's Health Initiative randomised trial of more than 68,000 women found that it did not.[17]
Sources
· 20 claims from 6 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
The Australian Bureau of Statistics defines correlation as a statistical measure, expressed as a number, of the size and direction of a relationship between two or more variables. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Correlation (retrieved 2026-11-14)
- [2]
A correlation between two variables does not automatically mean that a change in one causes the change in the other. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Correlation (retrieved 2026-11-14)
- [3]
The ABS gives the example of ice cream and sunscreen sales, which rise and fall together over the year because of the season, not because one affects the other. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Limitations (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Limitations (retrieved 2026-11-14)
- [4]
In practice it is much harder to establish cause and effect than to establish a correlation. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Causation (retrieved 2026-11-14)
- [5]
The ABS says a controlled study, in which a population is split into two groups comparable in almost every way that then receive different treatments, is the most effective way to establish causality. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Establishing causation (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Establishing causation (retrieved 2026-11-14)
- [6]
Ethical limits on controlled studies mean observational studies are often used instead to investigate correlation and causation, adding evidence to the wider process of deciding whether a causal link exists. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Establishing causation (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Establishing causation (retrieved 2026-11-14)
- [7]
Under the GRADE approach used by Cochrane, evidence from non-randomized studies should generally be rated down two levels because of the risk of confounding and selection bias. confirmedas of 2026-11-07
- Cochrane Handbook, Chapter 14: Completing 'Summary of findings' tables and grading the certainty of the evidence · Cochrane · Primary source · Section 14.2.1 (retrieved 2026-11-07)
- [8]
For two variables, correlation is measured by the correlation coefficient r, which runs from +1.0 to -1.0; a negative value means the variables move in opposite directions and a positive value means they move together. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Correlation coefficient (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Correlation coefficient (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Correlation coefficient (retrieved 2026-11-14)
- [9]
Correlation coefficients usually measure a linear relationship, so a real but non-linear relationship can produce a coefficient close to zero. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Limitations (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Limitations (retrieved 2026-11-14)
- [10]
The ABS says the correlation coefficient should not be used to say anything about cause and effect, and that causality is commonly misunderstood by people who assume a correlation means there must be a causal link. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Limitations (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Causation (retrieved 2026-11-14)
- [11]
Causation means one event is the result of the other, a cause-and-effect relationship; the ABS's example is that smoking causes an increased risk of lung cancer, whereas smoking is correlated with alcoholism but does not cause it. confirmedas of 2026-11-14
- Correlation and causation · Australian Bureau of Statistics · Primary source · Causation (retrieved 2026-11-14)
- Correlation and causation · Australian Bureau of Statistics · Primary source · Causation (retrieved 2026-11-14)
- [12]
The Cochrane Handbook says confounding occurs when there are common causes of both the choice of intervention and the outcome, so that the observed association differs from the causal effect. confirmedas of 2026-11-07
- Cochrane Handbook, Chapter 25: Assessing risk of bias in a non-randomized study · Cochrane · Primary source · Section 25.2.1 (retrieved 2026-11-07)
- [13]
In an observational study, researchers observe people or measure outcomes without trying to affect the outcome, for example without giving a treatment. confirmedas of 2026-11-07
- NCI Dictionary of Cancer Terms: observational study · National Cancer Institute (NIH) · Primary source · definition (retrieved 2026-11-07)
- [14]
GRADE allows certainty in non-randomized evidence to be rated up for a large effect, a dose-response gradient, or when plausible confounding would only weaken the observed effect. confirmedas of 2026-11-07
- Cochrane Handbook, Chapter 14: Completing 'Summary of findings' tables and grading the certainty of the evidence · Cochrane · Primary source · Key Points (retrieved 2026-11-07)
- Cochrane Handbook, Chapter 14: Completing 'Summary of findings' tables and grading the certainty of the evidence · Cochrane · Primary source · Table 14.2.a (retrieved 2026-11-07)
- [15]
A cross-sectional study collects information at a single point or over a short period; it can suggest an association between a risk factor and a disease but does not prove they are linked. confirmedas of 2026-11-07
- NCI Dictionary of Cancer Terms: cross sectional study · National Cancer Institute · Primary source · definition (retrieved 2026-11-07)
- NCI Dictionary of Cancer Terms: cross sectional study · National Cancer Institute · Primary source · definition (retrieved 2026-11-07)
- [16]
NCI's PDQ evidence ranking puts randomized controlled trials first, then nonrandomized controlled trials, then cohort or case-control studies, then ecologic and descriptive studies, and expert opinion last. confirmedas of 2026-11-07
- Levels of Evidence for Cancer Screening and Prevention Studies (PDQ) · National Cancer Institute · Primary source · Evaluation of Evidence, Study Design (retrieved 2026-11-07)
- Levels of Evidence for Cancer Screening and Prevention Studies (PDQ) · National Cancer Institute · Primary source · Evaluation of Evidence, Study Design (retrieved 2026-11-07)
- [17]
The Women's Health Initiative ran a randomized trial of more than 68,000 postmenopausal women alongside an observational study of more than 93,000, and found that hormone replacement therapy did not prevent heart disease as had once been thought. confirmedas of 2026-11-07
- Women's Health Initiative (WHI) · National Heart, Lung, and Blood Institute · Primary source · At a Glance (retrieved 2026-11-07)
- Women's Health Initiative (WHI) · National Heart, Lung, and Blood Institute · Primary source · At a Glance (retrieved 2026-11-07)
- Women's Health Initiative (WHI) · National Heart, Lung, and Blood Institute · Primary source · At a Glance (retrieved 2026-11-07)
- [18]
Two randomized trials (ATBC and CARET) found that heavy smokers given beta-carotene supplements had higher lung cancer incidence and mortality, and CARET was stopped early for harm. confirmedas of 2026-11-07
- Lung Cancer Prevention (PDQ) - Health Professional Version · National Cancer Institute · Primary source · Evidence summary, beta-carotene in smokers (retrieved 2026-11-07)
- Lung Cancer Prevention (PDQ) - Health Professional Version · National Cancer Institute · Primary source · Beta-carotene, CARET (retrieved 2026-11-07)
- [19]
The Cochrane Handbook notes that a fall in risk from 4% to 3% can be described either as a 25% relative reduction or as a 1 percentage point absolute reduction, and that clinicians may be more inclined to prescribe a treatment described the first way. confirmedas of 2026-11-07
- Cochrane Handbook, Chapter 15: Interpreting results and drawing conclusions · Cochrane · Primary source · Section 15.4.1 (retrieved 2026-11-07)
- Cochrane Handbook, Chapter 15: Interpreting results and drawing conclusions · Cochrane · Primary source · Section 15.4.1 (retrieved 2026-11-07)
- [20]
The ASA's fifth and sixth principles are that a p-value, or statistical significance, does not measure the size of an effect or the importance of a result, and that a p-value by itself is not a good measure of evidence about a model or hypothesis. confirmedas of 2026-11-14
- American Statistical Association releases statement on statistical significance and p-values · American Statistical Association · Primary source · 2016-03-07 · principle 5 (retrieved 2026-11-14)
- American Statistical Association releases statement on statistical significance and p-values · American Statistical Association · Primary source · 2016-03-07 · principle 6 (retrieved 2026-11-14)
Revision history (1)
- Page created.
Created Nov 14, 2026. Last reviewed by an editor on Nov 14, 2026. Next scheduled review: Nov 14, 2027.
Cite this page
"Correlation vs causation: why linked doesn't mean caused." ContentLora, updated Nov 14, 2026. https://contentlora.com/explain/correlation-vs-causation-explained
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.