Correlation vs Causation
Also known as: spurious correlation, confounding variable, causal inference, correlation coefficient, causality
Correlation describes a statistical relationship between two variables -- when one tends to increase or decrease as the other does.
Causation describes a direct mechanism by which one variable produces changes in another. These are distinct: ice cream sales and drowning rates both increase in summer (correlated), but eating ice cream does not cause drowning -- the real cause of both is hot weather (a confounding variable). Confusing correlation with causation is one of the most common analytical errors in business decisions, public health claims, and media coverage of research. Establishing causation requires controlled experiments, natural experiments, or specific statistical methods beyond simple correlation analysis.
Correlation vs Causation, explained at three levels
At home
Correlation means two things tend to happen together -- when one goes up, the other tends to go up (or down) too. Causation means one thing actually causes the other to happen. These are not the same: countries with more TVs per household also tend to have higher life expectancy (correlated), but TVs don't make people live longer -- both are caused by higher wealth. A classic example: people who carry lighters are more likely to get lung cancer than people who don't. Lighters don't cause lung cancer -- smoking does, and smokers carry lighters. The lighter correlates with cancer, but doesn't cause it.
At work
Confusing correlation with causation in business leads to wasted money and poor decisions. If you notice that customers who view your product video have higher conversion rates, the correlation is real -- but is viewing the video causing higher conversion, or are highly motivated buyers watching the video because they were already likely to buy? If the second, sending the video to everyone won't increase conversions -- you'd be targeting a symptom (watching the video) rather than a cause. Distinguishing correlation from causation in business analysis prevents spending resources on interventions that don't actually change outcomes.
At home
Confounding variables are the primary mechanism producing spurious correlations: a third variable that causes both the correlated variables. Height correlates with salary (taller people earn more on average), but height doesn't cause higher salary -- both are influenced by childhood nutrition, healthcare access, and socioeconomic status, which also affect career opportunities. Identifying potential confounders requires asking: what else could explain why these two things move together? Randomized controlled experiments (randomly assigning people to treatment and control groups) are the most reliable method to eliminate confounders, because random assignment distributes both known and unknown confounders equally between groups.
At work
Methods for establishing causation without randomized experiments (observational studies): natural experiments (a policy change or external event randomly affects some people but not others, creating a quasi-experimental comparison); difference-in-differences analysis (comparing before-and-after changes in a treated group to before-and-after changes in an untreated group); instrumental variables (using a variable that affects the treatment but has no direct effect on the outcome to isolate causal effects); and regression discontinuity (comparing outcomes just above and below an arbitrary threshold, like a policy cutoff score). These methods are used when experiments are impossible or unethical.
At home
Judea Pearl's causal hierarchy (the 'ladder of causation'): observation ('seeing') answers questions like 'what correlates with what?'; intervention ('doing') answers 'what happens if I change X?'; counterfactual ('imagining') answers 'what would have happened if X had been different?'. Traditional statistics operates primarily at the observation level. Causal inference methods attempt to answer intervention and counterfactual questions from observational data using causal graphs (DAGs -- directed acyclic graphs) that represent assumed causal mechanisms. This framework makes the assumptions required to infer causation from observational data explicit and testable.
At work
Selection bias as a causation-confusion mechanism in business analysis: customers who contact support have higher churn rates than those who don't (correlated). Does contacting support cause churn? Or do customers who are already about to churn (for other reasons) contact support more? If it's the second, improving support quality won't reduce churn -- you'd be addressing a symptom of a different problem. Randomized assignment of support contact (experimentally assigning some at-risk customers to receive proactive support outreach) is the most reliable test. Selection bias -- the tendency for the people or events you observe to be systematically different from the full population -- is a pervasive source of spurious causal conclusions in business analysis.
Real-world examples
At home
Sophie read a news article claiming that 'people who make their beds in the morning are more successful.' The article cited a survey showing bed-makers earned higher incomes. She recognized this as a correlation, not causation: the same personality traits, childhood discipline, or home environment that lead people to make their beds also likely contribute to professional success. Making your bed doesn't cause higher income; both are likely expressions of underlying traits or circumstances. She was skeptical of productivity advice built on this kind of correlation research.
At work
Thomas managed a retail marketing team. An analyst found that customers who received email newsletters had a 40% higher repeat purchase rate than customers who didn't. The analyst recommended increasing the email list aggressively. Thomas recognized a potential correlation-causation issue: customers who opted into the email list were already more engaged with the brand (that's why they opted in) -- they might have had higher repeat purchase rates regardless. He ran an experiment: randomly selecting from the opted-in list and only sending emails to half of them, comparing repeat purchase rates. The email-sent group showed only a 5% improvement, while the opted-in-but-not-sent group had nearly the same high rate. Most of the difference was selection bias, not email effectiveness.
Common confusions
- Many people assume that a very strong correlation (close to 1.0) must indicate causation — Correlation strength (how closely two variables move together) is separate from whether one causes the other. A correlation of 0.99 can be entirely spurious -- driven by a third variable or coincidence in time trends. Nicolas Cage movies release per year correlates 0.87 with swimming pool drownings (both trend over time with population and entertainment spending trends). Very high correlations between unrelated variables appear regularly when examining large datasets with many variables. Strength of correlation is not evidence of causation.
- Some assume that because correlation does not imply causation, correlational findings are worthless — Correlational findings are valuable even without establishing causation: they identify relationships that are worth investigating further, support predictions (if A and B are correlated, knowing A helps predict B even without a causal mechanism), and generate hypotheses for experimental testing. Most of medical science's most important findings (smoking and cancer, cholesterol and heart disease, diet and diabetes) were initially established through correlational studies before causal mechanisms were confirmed. Correlation is not causation but it is useful information -- it tells you where to look for causal relationships.
For analysts
When reviewing client analysis or data presentations, the correlation-causation question is one of the highest-value checks. The standard questions: could there be a confounding variable that explains this relationship? Are the groups being compared actually comparable (or is selection bias creating apparent differences)? Has anyone tested this with an experiment? For marketing analytics specifically, the self-selection problem is pervasive: customers who engage with your content, products, or loyalty programs were already more likely to be engaged -- demonstrating causation requires experimental intervention, not just observing that engaged customers have better outcomes.
History & origin
The logical fallacy of inferring causation from correlation has been documented since ancient Greek philosophy (post hoc ergo propter hoc -- 'after this, therefore because of this'). Karl Pearson and Francis Galton developed the mathematical correlation coefficient in the 1880s-1890s. Pearson himself warned that correlation could not establish causation. In practical research settings, the warning is widely understood but frequently violated in media coverage, business analysis, and policy arguments. Tyler Vigen's 'Spurious Correlations' website (2014, showing that per capita cheese consumption correlates highly with deaths by bedsheet-tangling) popularized the issue humorously. Judea Pearl's work on causal inference ('The Book of Why', 2018) developed rigorous mathematical frameworks for determining causal relationships from observational data -- bringing causality from philosophy into practical statistics.