How Correlation Differs From Causation

Ice cream sales and drowning deaths rise and fall together every summer, tracking each other so closely that a scatter plot of the two would look almost like a straight line. Nobody buying a cone is drowning anyone. Both climb because hot weather drives people toward pools, lakes, and beaches at the same time it drives them toward ice cream trucks. That shared driver, not any real link between the two variables themselves, is what statisticians call a confounder, and confounding is the single most common reason a correlation gets mistaken for a cause. This article walks through what correlation can and cannot tell you, the specific ways it misleads, and the actual methods researchers use to establish that one thing causes another.

What correlation actually measures

Correlation — a statistical measure of how tightly two variables move together — is expressed as a number between -1 and 1. A correlation of 0.9 between two variables means they track each other closely across a dataset; it says nothing about which one, if either, is driving the other, or whether some third factor is driving both. A correlation near 0 means the two variables show no consistent linear pattern together, but that alone does not rule out a real relationship that simply isn't linear.

The most widely used version, the Pearson correlation coefficient, only measures linear association: how well the data points fit a straight line. That is a narrower tool than most people assume. A handful of extreme outlier values can drag the coefficient up or down dramatically, and a genuinely strong relationship that curves — like the U-shaped relationship between age and reaction time, which improves through childhood and worsens after middle age — can produce a Pearson correlation near zero even though the two variables are clearly connected.

Causation requires something correlation cannot supply on its own: evidence that changing one variable actually produces a change in the other, holding everything else constant. That's a much higher evidentiary bar. It usually demands either a controlled experiment, where researchers manipulate one variable directly, or a natural experiment, where circumstances outside anyone's control approximate that manipulation closely enough to rule out alternative explanations one by one.

Think of correlation as a smoke detector and causation as the fire investigation report. The detector tells you something is going on and where to look. It does not tell you whether the smoke came from a kitchen fire, a faulty appliance, or someone burning toast. Treating the alarm itself as proof of the fire's cause is exactly the mistake correlation-causation confusion makes at scale, across news coverage, business decisions, and public policy.

A second way to picture it: correlation is like two people arriving at the same party at the same time. Maybe one drove the other. Maybe they carpooled with a third person. Maybe they just happen to both finish work at 6 p.m. and live nearby. Arriving together tells you almost nothing about who, if anyone, caused whom to be there.

The three ways correlation misleads

Three distinct traps hide behind the phrase "the data shows a relationship," and they get confused with each other constantly, both in casual conversation and in headlines that should know better.

• Confounding: A third variable causes both of the observed variables, creating a correlation between them with no direct link at all. Cities with more firefighters on staff also tend to have more total fire damage, not because firefighters cause fires but because larger, more fire-prone cities both employ more firefighters and generate more damage. The city's size and building density is the confounder driving both numbers.
• Reverse causation: The arrow of influence runs backward from what people intuitively assume. Companies with more customer complaints often also have more support staff, and it is tempting to conclude the staff are somehow generating complaints. In reality, complaint volume typically drives the hiring decision — the causal arrow points the opposite direction from the naive reading.
• Coincidence in small or cherry-picked data: With enough variables scanned against each other, some will correlate strongly by chance alone, with no plausible mechanism connecting them. This is sometimes called a spurious correlation, and it becomes more likely the more comparisons a researcher or algorithm runs without a prior hypothesis to test.

Tyler Vigen, a data analyst who launched the "Spurious Correlations" project while a student at Harvard Law School, used custom software to scan large public datasets and catalog dozens of statistically strong (r above 0.9) but causally meaningless pairings — most famously, the number of films Nicolas Cage appeared in each year correlating tightly with the number of U.S. swimming pool drownings over the same decade. The project, later published as a 2015 book, remains one of the most cited teaching examples for the difference between a strong statistical relationship and an actual cause.

Confounders in the wild: a closer look

Confounding is worth slowing down on because it's the mechanism behind most real-world correlation-causation mistakes, not just the ice cream and drowning example everyone learns in an introductory statistics class. A confounder has to satisfy two conditions: it must be associated with the variable being treated as the "cause," and it must independently influence the outcome. When both conditions hold, a correlation appears between the two original variables even if neither one touches the other directly.

Consider a few well-documented patterns that follow this same structure. Coffee drinking has repeatedly shown up as correlated with lower rates of certain diseases in observational studies, but coffee drinkers in many of these populations also tend to smoke less, exercise more consistently, and have higher average income, any of which could independently affect health outcomes. Researchers try to statistically adjust for these known confounders, but an unmeasured or unknown confounder can always remain, which is exactly why observational nutrition and lifestyle research is treated with more caution than randomized trial results.

The table below lays out several classic confounding examples side by side, showing the surface-level correlation, the variable people mistakenly treat as the cause, and the actual hidden driver connecting both.

Observed Correlation	Naive (Wrong) Interpretation	Actual Confounder
Ice cream sales rise with drowning deaths	Ice cream consumption causes drowning risk	Hot summer weather drives both swimming and ice cream demand
Cities with more firefighters have more fire damage	Firefighters cause worse fires	City size and building density drive both firefighter staffing and total damage
Shoe size correlates with reading ability in children	Bigger feet improve reading skill	Age drives both — older children have bigger feet and stronger reading skills
Countries with more chocolate consumption report more Nobel laureates	Chocolate boosts intelligence	Wealthier countries have both higher chocolate consumption and better-funded research institutions
Sunscreen use correlates with higher skin cancer rates in some surveys	Sunscreen causes skin cancer	People with high sun exposure, the real risk factor, are also more likely to apply sunscreen

Every row shows a statistically real relationship between the two listed variables, but the true causal driver sits outside the pair entirely.

The pattern across every row in that table is identical: two variables move together because a third, often unmeasured, factor is quietly driving both. Spotting a confounder usually starts with a simple question — is there any factor that could plausibly cause both things I'm comparing, rather than one causing the other? That single question catches a large share of the correlation-causation errors that circulate in casual reporting.

How researchers actually establish causation

Randomized controlled trials (RCTs) are considered the gold standard for establishing causation because random assignment spreads confounders evenly across groups by design, not by hope or statistical adjustment after the fact. If a drug trial randomly assigns 500 patients to receive the drug and 500 to receive a placebo, age, diet, genetics, income, and every other lurking variable should, on average, distribute roughly the same way across both groups purely by chance. That leaves the treatment itself as the one systematic difference between the groups, which is what allows researchers to attribute any difference in outcomes to the treatment rather than to some hidden factor.

This is why RCTs work even when researchers don't know every possible confounder in advance. You don't need to identify and measure every lurking variable if randomization has already balanced it across your groups statistically. That's a genuinely powerful property, and it's the main reason drug regulators like the FDA require randomized trial evidence, not just observational correlation, before approving most new treatments.

When randomization isn't ethically or practically possible — nobody can randomly assign people to smoke cigarettes for decades to study lung cancer risk — researchers lean on a different toolkit. Natural experiments look for situations where an external event, like a policy change or a randomly timed disaster, effectively assigns people to different conditions without a researcher doing the assigning. Instrumental variable analysis finds a factor that influences the suspected cause but has no direct effect on the outcome except through that cause, which lets statisticians tease out the causal portion of an observed relationship. And when neither of those is available, researchers turn to a structured framework for weighing observational evidence.

What are the Bradford Hill criteria?

The Bradford Hill criteria are nine considerations proposed by epidemiologist Sir Austin Bradford Hill in a 1965 paper to help judge whether an observed association is likely causal. They include the strength of the association, consistency across independent studies, a biological gradient or dose-response relationship, and biological plausibility, among others. Hill developed the framework while working on the research establishing the link between smoking and lung cancer.

None of the nine Bradford Hill criteria functions as a strict checklist where every box needs a checkmark. They're better understood as viewpoints that, taken together, make a causal explanation more or less credible than a coincidental one. The strength criterion looks at how large the association is; smokers historically showed lung cancer death rates roughly 10 to 30 times higher than nonsmokers, a magnitude far beyond what confounding alone typically produces. The biological gradient criterion asks whether more exposure produces more effect, and cigarette research consistently showed lung cancer risk rising with the number of cigarettes smoked per day. Consistency asks whether independent studies, using different methods and different populations, keep arriving at the same conclusion.

Does a strong correlation ever imply causation on its own?

A strong correlation raises the probability that a causal link exists somewhere in the picture, but the coefficient itself never confirms which variable, if either, is doing the causing. The strength of a correlation describes how tightly two variables move together, not why they move together. A correlation of 0.95 built on a genuine causal mechanism and a correlation of 0.95 built entirely on shared confounding look statistically identical on a scatter plot — the number alone cannot distinguish them.

Can causation exist without a strong measurable correlation?

Yes. A real causal effect can be masked by a weak or nonexistent linear correlation if the relationship is nonlinear, if the effect only appears in a subgroup of the population, or if a separate factor is suppressing the visible association in the aggregate data. This is one reason researchers combine correlation analysis with subgroup analysis and mechanistic study rather than relying on a single coefficient.

Why this distinction matters outside the classroom

News headlines routinely compress "researchers found an association" into "X causes Y," and that compression changes what a reasonable person should actually do with the information. A headline claiming coffee drinkers live longer, based on an observational study that simply tracked coffee habits and mortality over time, invites a very different response than a randomized trial demonstrating the same result experimentally. The observational version may just reflect that healthier, wealthier, less-stressed people happen to drink more coffee in the first place, with coffee itself doing nothing.

Public policy runs into this trap constantly and at much higher stakes than a coffee headline. A city that installs surveillance cameras and later observes falling crime rates cannot honestly claim the cameras caused the drop without first ruling out simultaneous changes in policing strategy, local economic conditions, demographic shifts, or a broader regional crime trend that had nothing to do with the cameras. Treating correlation as causation in a policy decision can direct scarce public resources toward the wrong lever entirely, while the actual driver of the outcome goes unaddressed.

The same trap shows up regularly in business analytics. A company might notice that customers who use a particular feature churn less often and conclude the feature reduces churn, then invest heavily in pushing more users toward it. But if engaged, satisfied customers were already more likely to explore that feature in the first place, the causal arrow may run from satisfaction to feature use rather than the other way around, and the investment produces far less impact than the correlation implied. Data teams that skip causal reasoning and act purely on correlation risk optimizing for a signal instead of the outcome they actually care about.

Medicine offers perhaps the highest-stakes version of this problem. Hormone replacement therapy was linked to lower heart disease risk in large observational studies through the 1980s and 1990s, a correlation strong enough that it shaped clinical guidance for years. When the Women's Health Initiative ran a large randomized controlled trial to test the claim directly, the results showed increased cardiovascular risk in some subgroups, not the protective effect the observational data had suggested. Women who chose hormone therapy in the earlier observational studies tended to be healthier and more health-conscious to begin with, a confounding pattern now referred to as the "healthy user bias," and it had inflated the apparent benefit for years before the randomized evidence corrected course.

Common mistakes people make even after learning the difference

Understanding that correlation isn't causation in the abstract doesn't automatically prevent someone from making the mistake in a specific case, and a few recurring errors are worth naming directly.

• Assuming the more intuitive causal direction is automatically the correct one. When two variables correlate, people default to whichever causal story feels more natural rather than checking whether the reverse direction, or a shared third cause, fits the data just as well.
• Treating "statistically significant" as equivalent to "causally proven." A significant correlation coefficient tells you the relationship is unlikely to be pure random noise in that sample; it says nothing about mechanism, and confounded relationships can be just as statistically significant as genuinely causal ones.
• Dismissing an entire finding because it's "just correlation." Correlational evidence isn't worthless. It's often the first and only signal available before an expensive experiment is feasible, and a well-designed observational study that carefully adjusts for known confounders can still meaningfully raise or lower confidence in a causal hypothesis.
• Expecting a single study to settle the question. Even a well-run randomized trial can produce a fluke result. Bradford Hill's consistency criterion exists precisely because a causal claim becomes credible through replication across multiple independent studies and methods, not through any single result, however strong.
• Ignoring effect size in favor of statistical significance. A correlation can be statistically significant and causally real while still being far too small to matter practically — this connects directly to broader confusion around statistical significance covered in why P values are misunderstood.

The takeaway

Correlation tells you two things moved together; causation tells you why. Confusing the two is not a minor technicality — it's the gap between a headline that sounds convincing and a claim that has actually been tested against alternative explanations. Randomized controlled trials remain the strongest tool for closing that gap because they neutralize confounders by design rather than by statistical adjustment after the fact, which is why they carry more weight than any observational correlation, however strong. The next time a chart shows two lines rising in tandem, the useful question isn't "are they related?" It's "what would have to be true for one to be causing the other, and has anyone actually checked?"

---

Read next:

• Why P Values Are Misunderstood
• Why Randomized Trials Are Powerful
• How Survivorship Bias Hides Failures

FAQ

Q: What is the difference between correlation and causation?

Correlation means two variables move together in a statistical pattern, measured by a coefficient between -1 and 1. Causation means changing one variable actually produces a change in the other. Correlation can be evidence toward causation but is never sufficient proof by itself, since confounding, reverse causation, or coincidence can all produce a correlation with no direct causal link at all.

Q: Can two things be strongly correlated with no causal relationship at all?

Yes. A confounding variable can independently drive both, the direction of causation can run backward from what seems intuitive, or the relationship can be a coincidence produced by scanning a small or cherry-picked dataset. Tyler Vigen's "Spurious Correlations" project documents dozens of real, strongly correlated but causally meaningless data pairings drawn from public U.S. datasets.

Q: Why do randomized controlled trials solve the correlation problem that observational studies can't?

Random assignment spreads confounding factors evenly across the treatment and control groups purely by chance, so researchers don't need to identify or measure every lurking variable in advance. Any remaining systematic difference in outcomes between groups is far more likely to reflect the treatment itself than a hidden third variable, which is why regulators require RCT evidence before approving most new medical treatments.

Q: What are the Bradford Hill criteria used for?

The Bradford Hill criteria are nine considerations, proposed by epidemiologist Austin Bradford Hill in 1965, used to judge whether an observed association in non-randomized data is likely to be causal. They include strength of association, consistency across independent studies, dose-response relationship, and biological plausibility, and they were developed largely through research establishing the link between smoking and lung cancer.

Q: How can I tell if a news report is confusing correlation with causation?

Check whether the underlying study was observational, meaning researchers simply measured existing data without intervening, or experimental, meaning researchers actively randomized participants into groups. Observational studies support correlation claims and can raise or lower confidence in a causal hypothesis, but only well-designed experiments, or a large body of converging observational evidence assessed against a framework like the Bradford Hill criteria, support a firm causal claim.

Q: Does correlation always mean one variable influences the other?

No. A correlation can appear even when neither variable influences the other at all, if both are independently driven by a third factor, such as a confounder, or if the pattern emerged by chance from scanning enough variables against each other. Establishing which explanation applies requires additional evidence beyond the correlation coefficient itself, such as a controlled experiment or a structured causal framework.