Correlation vs Causation
Correlation is when two things move together. Causation is when changing one thing makes the other change. The two are easy to confuse, and the confusion is everywhere: in headlines about coffee and cancer, screen time and attention spans, and money and happiness. A strong correlation, even a perfect one, is still only an association until an experiment shows the cause.
This guide explains the difference in plain language, works through real numbers with the site’s Correlation Coefficient Calculator, and gives you a checklist you can use on any causal claim you meet.
Short answer: correlation means two variables move together, measured by the correlation coefficient r, which runs from β1 to +1. Causation means changing one variable directly produces a change in the other. Only a controlled experiment β changing one thing while holding everything else constant β can prove causation.
What Is Correlation?
Correlation describes how two variables move together. When one goes up, does the other tend to go up too β or down, or neither? If taller children tend to be heavier, height and weight are correlated. If exam revision time and exam marks rise together, those two are correlated as well. The word for the general idea is association: two variables are associated when their values are linked in a regular way.
The standard measure of correlation is Pearson’s correlation coefficient, r. It runs from β1 to +1. An r of +1 means a perfect positive straight-line relationship: every point sits exactly on a rising line. An r of β1 means a perfect negative one. An r near 0 means no straight-line relationship at all. Most real data falls in between β an r of 0.5 is moderate, an r of 0.9 is strong. The NIST/SEMATECH e-Handbook describes the correlation coefficient as a measure of the linear relationship between two variables, and r is built from the standard deviations of the two variables, so it is a pure number with no units.
Three facts about r matter for everything that follows. First, r measures linear association only. A perfect curve β say, y = xΒ² β can have an r near 0 even though the two variables are perfectly linked, because the link is not a straight line. Second, r = 0 rules out a linear relationship, not all relationships. And third, r is symmetric: the correlation between A and B is the same as the correlation between B and A. That symmetry is exactly why r can never, on its own, tell you which variable causes the other β or whether either does.
What Is Causation?
Causation is a much stronger claim. To say that A causes B means that changing A produces a change in B, when everything else is held constant. Push the dose up and the effect goes up; remove the cause and the effect disappears. A classic formulation is the intervention test: if you could intervene on A alone, leaving the rest of the world untouched, would B change? If yes, A causes B. If not, the link between them is something else.
That definition has two consequences. The first is direction: causation has a direction, from cause to effect, while correlation does not. The second is that a causal claim needs a mechanism β a plausible chain of steps from A to B. “More ice cream is followed by more drownings” has no mechanism: eating ice cream does not make you drown. “Smoking causes lung cancer” does have one: smoke damages lung tissue, and damaged tissue can turn cancerous. A correlation with no plausible mechanism should always raise suspicion.
Here is the asymmetry that the whole article turns on: causation implies correlation, but correlation does not imply causation. If A genuinely causes B, the two will move together and r will pick that up. But the reverse is not true. A high r can arise in at least four different ways, and only one of them is direct causation β which is where the next section comes in.
Strong correlation (r = 0.96)
Spurious correlation (r = 0.999)
Variable A
Ice-cream sales
Variable B
Drownings
Direct cause: A causes B
Shared cause: C causes both
A
B
C
A
B
The Four Explanations for Any Correlation
Whenever two variables move together, there are only four possible stories that can explain it. The correlation coefficient cannot tell you which one applies β r is the same number in all four cases. Your job, when you read a headline or a study, is to work out which story the evidence actually supports.
1. Direct cause: A causes B
This is the story headlines usually assume. Rain causes umbrella sales to rise; turning a heater up causes the temperature to rise. The strongest case is smoking and lung cancer: smokers develop lung cancer far more often than non-smokers, the risk rises with the number of cigarettes smoked, and there is a clear mechanism β smoke damages lung tissue. When a correlation has a mechanism and survives an experiment, direct causation is the right conclusion. The problem is that people jump to this explanation first, before checking the other three.
2. Reverse causation: B causes A
Sometimes the arrow points the other way. Areas with more hospitals have more sick people β but hospitals do not make people sick; sick people go to hospitals. The same trap appears with depression and social-media use: heavy social-media users report more depressive symptoms, and it is often claimed that social media causes depression. But the arrow may run the other way β people who feel low may spend more time scrolling and withdrawing from face-to-face contact β or both directions may operate at once. A correlation never tells you which way the arrow points; only a study that changes one variable and watches the other can do that.
3. Confounding: a third variable causes both
This is the most common explanation of all. Ice-cream sales and drownings rise together through the summer β but ice cream does not cause drowning. Temperature causes both: hot weather sells ice cream and sends more people swimming. Temperature is a confounding variable (also called a lurking variable). You can try the ice-cream numbers in the Correlation Coefficient Calculator: the correlation is extremely strong, yet the cause is the weather, not the ice cream. A famous medical example is the claim that children who sleep with the light on become short-sighted. Later work showed that short-sighted parents both leave lights on and pass short-sightedness on to their children β the parents’ eyesight was the confounder.
4. Coincidence: pure chance
Sometimes two things line up for no reason at all. The best-known example is the number of films Nicolas Cage appeared in each year and the number of people who drowned by falling into a swimming pool: the two series track each other remarkably closely, yet no mechanism could possibly connect them. With thousands of pairs of variables to compare, some pairs will line up by chance alone. A single striking correlation, with no mechanism and no experiment behind it, is weak evidence β it is a coincidence until proven otherwise.
| Explanation | What it means | Example |
|---|---|---|
| Direct cause | A produces a change in B | Smoking and lung cancer |
| Reverse causation | B produces a change in A | Depression and social-media use |
| Confounding | A third variable causes both | Ice cream, drownings and temperature |
| Coincidence | No real link; chance alignment | Nicolas Cage films and pool drownings |
What Is a Confounding Variable?
A confounding variable is a third factor that is related to both the supposed cause and the effect, and that distorts the apparent relationship between them. Temperature confounds the ice-cream and drownings link; parents’ eyesight confounds the light-on and short-sightedness link. To spot one, ask: what else changes with both variables? Anything that could plausibly affect both sides is a suspect. Researchers deal with confounders by measuring them and adjusting for them, by comparing like with like, or β most reliably β by random assignment, which spreads every confounder, known and unknown, evenly between the groups being compared.
Worked Example: A Strong Correlation That Proves Nothing
Take ten pairs of numbers, x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] and y = [2, 3, 5, 4, 6, 8, 7, 9, 11, 10]. As x rises, y rises with it β the points sit close to a straight line. Run the pair through the site’s correlation calculator and here is what comes back:
| Measure | Value |
|---|---|
| Sample size (n) | 10 |
| Mean of x (xΜ) | 5.500 |
| Mean of y (Θ³) | 6.500 |
| Sxy | 82.500 |
| Correlation coefficient (r) | 0.964 |
| Coefficient of determination (rΒ²) | 0.929 |
| Regression slope | 0.964 |
| Regression intercept | 1.200 |
| t-statistic | 10.20 |
| p-value | < 0.001 |
| 95% confidence interval for r | 0.851 to 0.991 |
That is a strong, statistically significant correlation: r = 0.964, with 92.9% of the variation in y explained by the straight-line relationship, and a p-value below 0.001, meaning a correlation this strong would be very unlikely to appear by chance alone. The regression line is y = 1.200 + 0.964x, so each extra unit of x is associated with an extra 0.964 units of y. Now say it plainly: even this proves nothing about cause. The numbers say that x and y move together; they do not say that x moves y, that y moves x, or that a third factor is moving both. Without an experiment, all four explanations from the previous section are still on the table.
Worked Example: A Spurious Correlation
Now the classic case. The table below is an illustrative eight-month dataset: monthly ice-cream sales and monthly drowning deaths, both rising through the summer. Enter these pairs into the correlation calculator to see what the tool returns:
| Month | Ice-cream sales ($000s) | Drownings |
|---|---|---|
| May | 85 | 4 |
| June | 120 | 5 |
| July | 150 | 7 |
| August | 175 | 8 |
| September | 210 | 10 |
| October | 250 | 12 |
| November | 290 | 14 |
| December | 330 | 16 |
| Measure | Value (tool-computed) |
|---|---|
| Sample size (n) | 8 |
| Mean of x (xΜ) | 201.250 |
| Mean of y (Θ³) | 9.500 |
| Sxy | 2535.000 |
| Correlation coefficient (r) | 0.999 |
| Coefficient of determination (rΒ²) | 0.997 |
| Regression slope | 0.050 |
| Regression intercept | β0.635 |
| t-statistic | 47.66 |
| p-value | < 0.001 |
| 95% confidence interval for r | 0.992 to 1.000 |
The tool-computed r is 0.999 β about as strong as a correlation gets, and highly statistically significant (p < 0.001). And it means nothing at all as a causal claim. Ice cream does not cause drowning. Both rise because the weather gets warmer: heat sells ice cream and sends more people swimming. The confounder is temperature (and the summer season). If you only looked at the number β 0.999, t = 47.66 β you would be more certain of a false claim than of the genuine one in the previous example. That is the whole lesson of this article in a single number.
What About a Perfect r = Β±1?
The calculator’s own presets make the point absolute. A perfect positive relationship gives r = 1.000 (regression slope 2, intercept 0); a perfect negative one gives r = β1.000 (slope β2, intercept 12). Every point sits exactly on the line β there is no scatter at all. And it still proves nothing about cause. As the Australian HSC curriculum puts it, even an r of +1 does not prove causation. Perfection in the number is perfection in the association, and association is all r ever measures.
Correlation vs Causation at a Glance
The two ideas sit side by side in almost every statistics course, and the table below is the fastest way to keep them apart.
| Aspect | Correlation | Causation |
|---|---|---|
| What it says | Two variables move together | Changing one variable changes the other |
| Measured by | The correlation coefficient r, from β1 to +1 | Experiments, especially randomised trials |
| Direction | None β correlation is symmetric | Has a direction, from cause to effect |
| Needs a mechanism? | No | Yes β a plausible chain from cause to effect |
| Proves the other? | No | A genuine cause implies correlation |
| If r = 0 | No straight-line relationship | Nothing β r = 0 says nothing about causation either way |
| Typical trap | Mistaking a big r for proof | Assuming one study settles it |
| Example | Ice-cream sales and drownings (r = 0.999) | Smoking and lung cancer |
Read the table from left to right and the lesson is plain: correlation is a description of data, causation is a claim about the world. A description can be measured in seconds with a calculator; a claim about the world needs evidence of the kind described in the next sections.
Does a High Correlation Prove Causation?
No. A high correlation β even a perfect one β only ever shows that two variables move together. The ice-cream example above had a correlation of r = 0.999 with a p-value below 0.001: the strongest possible statistical result, attached to a claim that is plainly false. Three quick questions expose most false causal claims: is there a plausible mechanism linking the two variables? Could a third variable explain the link? And could the arrow point the other way? If any of those questions has no good answer, the correlation is not yet evidence of causation, however large r is.
How Do You Prove Causation?
There is one way to establish causation properly, and several supporting checks that make a causal claim more or less believable. Use this checklist on any headline that says “X causes Y”:
1. Run a randomised controlled experiment. Randomly assign people or things to a treatment group and a control group, change only the treatment, and compare the outcomes. Random assignment is the key: it spreads every confounder, known and unknown, evenly between the groups, so the only systematic difference left is the treatment itself. If the groups then differ, the treatment caused the difference. This is the single method that can settle a causal question.
2. Check the mechanism. Is there a plausible step-by-step chain from cause to effect? Smoking damages lung tissue, which can turn cancerous. If no mechanism exists, be suspicious.
3. Check temporality. The cause must come before the effect. If the “effect” appears before the supposed cause, the claim is backwards.
4. Look for a dose-response. A bigger cause should produce a bigger effect: more cigarettes, higher cancer risk; a larger dose, a stronger response. A flat response weakens the claim.
5. Look for consistency. The same association should show up in different studies, different places and different populations β not in one study only. Read about how hypothesis testing works to see how studies weigh evidence like this.
6. Rule out confounders. List the third variables that could affect both sides, measure them, and adjust for them β or compare like with like. If the correlation survives, it is stronger evidence; if it disappears, the confounder was the story.
The Bradford Hill Criteria in Plain Language
In 1965, the statistician Sir Austin Bradford Hill set out nine considerations for judging whether an association is likely to be causal, in his paper “The Environment and Disease: Association or Causation?” They are a checklist for weighing evidence, not a verdict machine β no single criterion is decisive, and meeting all nine does not prove a cause. In plain language:
- Strength: a large association is harder to explain away by bias or confounding than a small one.
- Consistency: the same result appears in different studies, places and populations.
- Specificity: one cause linked to one effect is more convincing than a cause linked to everything.
- Temporality: the cause comes before the effect.
- Biological gradient (dose-response): more of the cause means more of the effect.
- Plausibility: a believable mechanism exists.
- Coherence: the claim fits with what is already known about the subject.
- Experiment: changing the cause changes the effect.
- Analogy: similar causes are known to produce similar effects.
Smoking and lung cancer scores well on almost all nine; ice cream and drowning scores well on almost none. That is the difference between a causal claim and a curiosity.
Common Mistakes
- r = 0.95 does not mean “95% likely causal”. r measures the strength of a straight-line association; it is not a probability that the link is causal. The ice-cream example had r = 0.999 and a 0% chance of being causal.
- “Statistically significant” does not mean “proven”. A p-value below 0.001 says the correlation is unlikely to be chance. It says nothing about which of the four explanations applies.
- Confusing correlation with the regression slope. r and the slope are different numbers: r is unit-free and always between β1 and +1, while the slope from a linear regression has units and can be any size. A slope of 0.964 and an r of 0.964 in the worked example are a coincidence of that data set, not the same thing.
- Cherry-picking the time window. Picking the years that make a pattern appear, and ignoring the years that break it, manufactures correlations.
- Ignoring a lurking variable. Forgetting to ask “what else changes with both variables?” is the single most common way to be fooled by a correlation.
Frequently Asked Questions
Tap a question to see the answer.
What is the difference between correlation and causation?
Correlation means two variables move together, measured by the correlation coefficient r, which runs from β1 to +1. Causation means changing one variable directly produces a change in the other. To prove causation, you need a controlled experiment: change one thing while holding everything else constant, and watch whether the other variable responds.
Does a high correlation prove causation?
No. A correlation, however high, only shows that two variables move together. The ice-cream and drownings example in this guide has a tool-computed r of 0.999 and a p-value below 0.001, yet ice cream plainly does not cause drowning β warm weather causes both. Even a perfect r of +1 or β1 is still only an association.
What is a confounding variable?
A confounding variable is a third factor that is related to both the supposed cause and the effect, and distorts their apparent relationship. Temperature confounds ice-cream sales and drownings; parents’ eyesight confounds the link between sleeping with the light on and short-sightedness. Researchers deal with confounders by measuring and adjusting for them, or by random assignment.
Does a negative correlation mean one thing causes the other to decrease?
No. A negative correlation only says the two variables move in opposite directions β when one goes up, the other tends to go down. It carries the same limits as a positive correlation: it cannot show direction or rule out a confounder. Only an experiment can tell you whether one variable causes the other to decrease.
Does a correlation of 0 mean there is no relationship?
Not necessarily. An r of 0 rules out a straight-line relationship only. Two variables can be perfectly linked by a curve β for example y = xΒ² β and still have a correlation near 0, because Pearson’s r measures linear association only.
How do you prove causation?
With a randomised controlled experiment: randomly assign subjects to a treatment group and a control group, change only the treatment, and compare outcomes. Random assignment balances all confounders, known and unknown. Supporting evidence includes a plausible mechanism, the cause coming before the effect, a dose-response, and consistent results across studies.
What is a spurious correlation?
A spurious correlation is an association with no causal link behind it β it is driven by a shared cause or by pure chance. Ice-cream sales and drownings are linked by the summer weather; the number of Nicolas Cage films and pool drownings line up by coincidence. A high r is no protection against a spurious result.
What is reverse causation?
Reverse causation is when the supposed effect actually causes the supposed cause. Areas with more hospitals have more sick people because sick people go to hospitals, not because hospitals make people sick. Similarly, heavier social-media use may follow from feeling low rather than cause it. Correlation cannot show which way the arrow points β only a study that changes one variable can.
Conclusion
Correlation tells you that two things move together; causation tells you that changing one changes the other. The number r, however large, can only ever give you the first of those. Use it to measure an association β r = 0.964 in a genuine case, r = 0.999 in a spurious one β then run the four explanations past it: direct cause, reverse causation, a confounder, or plain coincidence. And when a claim matters, insist on the experiment. That is the habit that separates reading a headline from understanding it.
Check Tools
Correlation Coefficient Calculator
Compute r, rΒ², the regression line and the significance test for any data set. Measure the correlation in your data
Linear Regression Calculator
Fit the line y = a + bx to your data, with the slope, the intercept and RΒ² explained. Fit a regression line to your data
Standard Deviation Calculator
Find the spread behind every correlation β r is built from the standard deviations of the two variables. Work out a standard deviation
Explore All Statistical Tools
Browse the full set of calculators and visualisers on the site. Browse all the statistical tools