Causation Is Not Correlation

Short answer

• Causation is not correlation because correlation only shows that two variables tend to move together, while causation means that a change in one variable produces a change in another. Two things can be strongly correlated even when neither causes the other. A third factor may drive both, the direction of influence may run the other way, or the pattern may be coincidence. A correlation is a useful clue, and it can point researchers toward a causal question, but on its own it does not prove that one thing causes another. To support a causal claim, you need evidence about design, direction, timing, and alternative explanations, not just a strong pattern.

Introduction

Headlines love causal language. A study finds that people who do one thing also tend to experience another, and the headline soon reads as if the first thing caused the second. Sometimes that reading is fair, because other evidence backs it up. Often it is not, because the study only measured an association and never tested whether changing one variable would change the other.

This distinction matters far beyond statistics classes. It shapes how you read health news, judge a product claim, interpret a graph shared online, and argue about policy. Mistaking correlation for causation can lead to useless advice and confident decisions built on a pattern that was never causal.

1. What Correlation Means

Correlation describes a relationship between two variables. A variable is anything that can take different values, such as daily temperature, hours of sleep, exam scores, or the number of cups of coffee a person drinks. When two variables are correlated, knowing something about one of them tells you something about the other, on average. They tend to move together in a predictable pattern.

A positive correlation means that as one variable rises, the other tends to rise as well. A negative correlation means that as one rises, the other tends to fall. If there is no correlation, the two variables show no consistent pattern of moving together.

One common summary of a straight line relationship is the correlation coefficient, written as r, which runs from negative 1 to positive 1. A value near 1 means a strong positive straight line pattern, near negative 1 a strong negative pattern, and near 0 little or no straight line pattern. StatsLearn stresses that r measures how closely points fall to a line, and says nothing on its own about why they line up.

2. What Causation Means

Causation is a stronger claim. To say that X causes Y is to say that a change in X produces a change in Y, through some mechanism, and that if X had been different, Y would have been different too, at least in the situation being studied. Causation has direction: it runs from cause to effect. It also has timing: the cause must come before the effect, or at least act through a process that unfolds in time.

Causation is rarely a matter of a single factor acting alone. Most outcomes have many contributing causes, and saying that something is a cause does not mean it is the only cause, or that it always produces the effect in every person. It means that, in the context studied, changing it would tend to change the outcome.

That idea of what would happen under a change is what separates causation from correlation. A correlation can be observed without changing anything: you simply record what already happened. A causal claim, by contrast, is about an intervention, real or imagined. If you lowered X, would Y fall? If you raised X, would Y rise? If you cannot answer that question with evidence, you do not yet have a causal finding, however striking the pattern looks.

Causal words such as causes, prevents, and triggers assert more than association words such as linked to and associated with. When a story upgrades an association into causal wording without new evidence, it has changed the claim, not just the style.

3. Four Ways a Correlation Can Arise

When two variables are correlated, there are four broad explanations to consider before accepting a direct causal story. Working through all four is the habit that protects you from the most common misreadings.

First, X may genuinely cause Y. This is the explanation people usually jump to, and sometimes it is correct. But it is only one of the possibilities, and it needs its own evidence.

Second, Y may cause X. This is reverse causation, where the arrow points the other way from the one first assumed. Section 5 works through examples.

Third, a third variable may cause both X and Y. This is confounding, and it is the explanation that fools even careful readers, because the pattern is real and often strong. Section 4 works through examples.

Fourth, the correlation may be coincidence. With enough variables and enough comparisons, some pairs will line up by chance, especially in short or cherry picked data. Section 6 works through examples.

4. Confounding: When a Third Factor Drives Both

Confounding is the most instructive of the four explanations, because the correlation involved is genuine, not a trick of small numbers. Two variables really do move together, but neither causes the other. Instead, a third variable, called a confounder or lurking variable, influences both of them, and that shared influence creates the appearance of a direct link.

The classic teaching example comes from Taft College statistics material published through LibreTexts. Ice cream sales and drowning deaths show a strong positive correlation, yet eating ice cream does not cause drowning. The third variable is outside temperature: when it gets hotter, more people buy ice cream and more people go swimming, which raises exposure to drowning risk. The correlation is real, but the causal story is false.

The same source gives a second example. Shoe size and spelling ability are consistently correlated among elementary school children, but neither causes the other. The confounder is age: older children spell better and are bigger, so they have larger shoes. Once age is taken into account, the apparent link fades. That is the signature of confounding: control for the third variable, and the correlation weakens or disappears.

A more consequential example comes from epidemiology teaching at Simon Fraser University. For years, observational studies suggested that hormone therapy protected the heart in women who chose to take it, but those women were wealthier, better educated, leaner, and more engaged with preventive care, all factors linked to lower heart disease risk. The randomized trial told a different story: according to the lesson, the 2002 trial result for combined estrogen and progestin was a hazard ratio of 1.29, with a 95 percent confidence interval from 1.02 to 1.63, for coronary heart disease, indicating harm rather than protection. This case warns that consistency across many observational studies does not remove confounding when the same kind of people keep selecting into the same group.

Confounding is hard to eliminate in observational research, because the list of possible third variables is very long. Researchers can adjust for known confounders, for example by comparing groups within the same age band, but they cannot adjust for a factor they never measured. The honest statement after adjustment is that the association survives after accounting for the factors measured, not that causation has been proven.

5. Reverse Causation: When the Arrow Points the Other Way

Reverse causation happens when the outcome drives the supposed cause, rather than the other way around. The correlation is real, but the direction is backwards. This error is common because the first causal story that comes to mind often feels natural, while the reversed story takes a moment longer to see.

StatsLearn gives a clear teaching example. Towns with more police officers tend to report more crime. It would be wrong to conclude that hiring officers manufactures crime. The more plausible direction is that heavy crime leads towns to hire more officers. Crime drives police numbers, not the reverse. Before locking in a direction, ask whether the response variable could be driving the explanatory variable instead.

The same source gives a second teaching example, presented as a classroom exercise rather than a study of real patients. Annual doctor visits and number of chronic conditions move together in the exercise data. Reading that pattern as proof that seeing the doctor causes chronic illness would reverse the arrow. People with more chronic conditions visit the doctor more often because they need care. The visits follow the illness; they do not create it.

Timing helps, but it does not settle direction on its own. A person may start a new habit because early symptoms appeared, so the habit looks like it precedes the illness while the illness process actually started first. To test direction, researchers look for designs where the supposed cause is assigned or changed independently of the outcome. Without that kind of design, direction remains an open question.

6. Coincidence and Spurious Correlations

Some correlations have no mechanism at all. They are coincidences, and they arise because the world contains a vast number of variables, and any two of them can happen to move together for a while. Science News Explores illustrates the point with two examples drawn from Tyler Vigen and his Spurious Correlations project. Over a ten year period, Americans fondness for margarine correlated strongly with the divorce rate in Maine, yet there is no reason to think one caused the other. Between 2000 and 2009, the more mozzarella cheese Americans ate, the more doctorates in civil engineering that United States universities awarded. Again, the two series moved together, and again there is no plausible mechanism connecting cheese consumption to engineering doctorates. These are coincidental patterns in unrelated data sets, not hidden causal stories waiting to be discovered.

Coincidence becomes more likely as data sets grow. If you compare thousands of variables, some pairs will correlate purely by chance, and searching only for striking pairs will find them. This is sometimes called data dredging or cherry picking: testing many relationships and reporting only the ones that pass a threshold. A correlation found this way is far weaker evidence than one predicted in advance and confirmed in fresh data.

Short series are especially vulnerable. Over a few years, two unrelated trends that both happen to rise, such as a growing population and a growing number of almost anything, will look correlated simply because both trend upward over time. That shared trend is a form of confounding by time, and it disappears once the trend itself is accounted for or once a longer, more varied series is examined.

The practical lesson is to ask how the pair was found. Was it predicted before the data were examined, on the basis of a plausible mechanism, or discovered by scanning many pairs until one looked impressive? Was it confirmed in a new data set, or does it rest on a single short series? A coincidence can produce a beautiful chart. It cannot produce a mechanism or a result that survives repetition.

7. How to Test a Causal Claim

A correlation is a starting point for a causal question, not an answer to it. The strongest evidence for causation comes from a well designed randomized comparative experiment: the researcher assigns the treatment, includes a comparison or control group, and uses random assignment to sort participants into groups, spreading confounding variables across the groups on average so they start out similar except for the treatment. When that holds, a difference in the outcome can be attributed to the treatment. Both the Taft College material and StatsLearn make this point: correlational analysis alone cannot demonstrate causality, while a properly designed experiment can, by ruling out other explanations.

Experiments are not always possible or ethical. Researchers cannot assign children to harmful exposures to see what happens, and they cannot randomize every policy or life choice. In those cases, causal reasoning relies on a combination of weaker evidence, handled with explicit caution. No single observational result should carry the full weight of a causal claim. Instead, look for several lines of evidence pointing the same way.

Use the following checklist when you meet a causal claim, whether in a news story, an advertisement, or a conversation.

One, what exactly is the claim? Separate the association from the causal wording. Does the source say linked, associated, or correlated, or does it say causes, prevents, boosts, or lowers? If the study only measured an association, causal wording is an upgrade that needs separate justification.

Two, what was the design? Was the supposed cause assigned by the researchers, or was it simply observed? Observational studies can reveal strong associations, but they cannot rule out lurking variables the way a randomized experiment can. If the design was observational, treat the causal claim as unproven until other evidence is added.

Three, what is the direction? Could the outcome drive the supposed cause instead? Check timing, and ask whether an early symptom, a prior condition, or a selection effect could have produced the pattern in reverse.

Four, what are the plausible confounders? Name at least one third variable that could drive both sides. If you can name a strong one that the study did not measure or control for, the claim is on shaky ground. If the study adjusted for the obvious confounders and the association survived, that strengthens the case, but it does not close it, because unmeasured confounders may remain.

Five, is there a plausible mechanism? Can you describe, even in general terms, how the cause would produce the effect? Coincidental correlations typically have none.

Six, does the pattern behave sensibly? Does more of the supposed cause go with more of the effect? Does the effect follow the cause in time? Does the finding repeat in different populations or studies? A single study is weaker than a pattern that survives repetition.

Seven, what would change your mind? A good causal claim implies that if the cause were removed or reversed, the effect would change. If no evidence is named that would count against the claim, you may be dealing with a slogan rather than a finding.

8. Common Misreadings to Avoid

Several misreadings recur so often that they are worth naming directly. Each one takes a genuine statistical idea and stretches it into a conclusion it cannot support.

Misreading one: a strong correlation proves causation. It does not. A correlation near 1 or near negative 1 only tells you that points fall close to a straight line. It cannot tell you whether one variable drives the other, whether a third variable drives both, or whether the pattern is coincidence. Strength of association and proof of causation are different questions.

Misreading two: statistical significance proves causation. Significance speaks to whether a pattern is unlikely to be chance under the study design. It says nothing about confounding, reverse causation, or whether the effect is large enough to care about. A significant association in an observational study is still an association.

Misreading three: after this, therefore because of this. Just because an outcome followed an event does not mean the event caused it. Without a comparison group, a before and after story cannot separate the effect of the event from everything else that changed at the same time.

Misreading four: a group pattern applies to every individual. Simon Fraser University teaching material names this the ecological fallacy: mistakenly inferring individual level relationships from group level data. A country level link between two averages need not describe any individual person. Keep the level of the data in mind, and do not transfer a group finding to a person without individual level evidence.

Misreading five: an overall pattern must hold within every subgroup. The reverse error is known as Simpson paradox, in which an association seen in aggregated data reverses when the data are split by a confounding variable. Whenever an overall result surprises you, ask whether a hidden grouping could be driving it.

Misreading six: controlling for one confounder settles the question. Adjustment only addresses the confounders measured and measured well. Residual confounding can remain when a factor is measured imprecisely or grouped too coarsely.

Misreading seven: correlation is useless. That overcorrects. Correlations generate hypotheses, reveal patterns worth investigating, and support prediction even when the mechanism is unknown. The error is not in using correlations, but in dressing them up as proven causes.

Frequently Asked Questions

Does correlation ever imply causation?

Not on its own. A correlation can motivate the experiments that test a link, but you can only conclude causation when the design rules out confounding, reverse causation, and coincidence. A randomized experiment does this most directly. Observational correlation, however strong, still leaves those alternatives open.

Can a correlation coefficient near 1 prove causation?

No. A value of r near 1 or near negative 1 only tells you that the points fall near a straight line. Judge the causal claim by the design and the supporting evidence, not by the size of r.

What is the difference between a confounding variable and a lurking variable?

A lurking variable is any variable outside your analysis that affects the results. It becomes a confounding variable when it is associated with both the explanatory variable and the response variable, giving an alternative explanation for the pattern.

What is reverse causation, in one sentence?

Reverse causation is when the supposed effect actually causes the supposed cause, so the causal arrow points the other way from the one first assumed, as when high crime leads towns to hire more police rather than police hiring causing crime.

Is an observational study worthless for causal questions?

No. Observational studies are often the only ethical or practical option, and they can build a strong case with a plausible mechanism, repetition across settings, sensible timing, and control of known confounders. What they cannot do alone is rule out unmeasured confounding, so their causal conclusions should state that limit.

Why do spurious correlations happen so often online?

Because charts are easy to make and striking coincidences are easy to share. With thousands of public data series available, anyone can find pairs that move together for a few years. Without a mechanism, a prior prediction, or confirmation in fresh data, such a chart is a curiosity, not a discovery.

How should I read a headline that says a study links two things?

Read linked as associated, not as caused. Then look for the design: was the factor assigned or merely observed, who was studied, and did the researchers control for obvious confounders? If the article does not answer those questions, treat the causal reading as unproven.

What is the single most useful question to ask about a causal claim?

Ask: what else could explain this pattern? If you can name a plausible confounder, a reversed direction, or a coincidence that fits the data just as well, the claim needs more evidence.

Sources & references

• Michelle Oja, Taft College, Research Methods for the Social and Behavioral Sciences, section 6.6.1 Correlation versus Causation, Social Sci LibreTexts. Opened and read 26 Oct 2026. Used for: correlation does not show causation, the need for a properly designed and controlled experiment, the ice cream sales and drowning example with temperature as the third variable, and the shoe size and spelling example with age as the confounder.
• Science News Explores, Explainer: Correlation, causation, coincidence and more, published 24 Jul 2015 and updated 25 Mar 2020. Opened and read 26 Oct 2026. Used for: the margarine and Maine divorce example, the mozzarella cheese and civil engineering doctorates example for 2000 to 2009, and the point that experiments use test and control groups while some exposures cannot ethically be assigned.
• StatsLearn, Correlation vs causation, with real examples, by Jude Wallis, updated 4 Sep 2026. Opened and read 26 Oct 2026. Used for: r as a number between negative 1 and 1, the three misleading mechanisms of confounding, reverse causation, and coincidence, the police officers and crime reverse causation example, the doctor visits teaching example, and the randomized comparative experiment account.
• Simon Fraser University, HSCI 230 Lesson 11: Confounding and Statistical Inference, developed by Dr. Kiffer G. Card, Faculty of Health Sciences. Opened and read 26 Oct 2026. Used for: the confounder definition as a common cause of exposure and outcome, the hormone therapy and Women's Health Initiative account including the 2002 trial hazard ratio of 1.29 with a 95 percent confidence interval from 1.02 to 1.63, and the definitions of Simpson paradox and the ecological fallacy.