What Are Correlation and Causation?
Karl Pearson introduced the product-moment correlation coefficient in 1895, building on work by Francis Galton. The formula compares how much two variables co-vary relative to how much each varies on its own. A coefficient near +1 means the variables track each other closely in the same direction. Near −1 means they track each other in opposite directions. Near 0 means little systematic linear relationship. The Pearson correlation guide covers the formula and interpretation in full detail.
The emphasis on intervention is intentional. Philosopher David Hume argued in 1748 that we never directly observe causation — we only observe sequences and regularities. Judea Pearl formalized this intuition in his work on causal inference, proposing that causation requires understanding what would happen if we intervened and set a variable to a specific value. That's the key distinction from correlation, which only describes observed patterns without any intervention.
Correlation describes a pattern in data. Causation explains that pattern with a mechanism. You need a mechanism — and evidence of one — before claiming cause and effect.
Correlation vs Causation: Side-by-Side
The table below covers the key points of difference. One note before reading it: causation always implies correlation in the underlying population (if A causes B, they will co-vary). But correlation does not imply causation — it only tells you that two variables move together, not why.
| Feature | Correlation | Causation |
|---|---|---|
| Definition | Statistical co-movement between two variables | One variable directly produces change in another |
| What it requires | Only observed data, no experiment needed | A mechanism and evidence of intervention effect |
| Can it be spurious? | Yes — can arise from confounders or chance | No — a true causal effect is real by definition |
| Primary tool | Correlation coefficient (Pearson r, Spearman ρ) | Randomized experiments, causal inference methods |
| Implies the other? | No — correlation does not imply causation | Yes — causation implies correlation in population |
| Policy use | Generate hypotheses, identify patterns | Justify interventions (drugs, policies, treatments) |
| Classic example | Ice cream sales and drowning rates | Smoking and lung cancer |
| Diagram | A ↔ B (associated, direction unknown) | A → B (A produces change in B) |
Four Reasons Correlation Can Appear Without Causation
Given any correlation between variables A and B, there are four distinct explanations to consider before concluding that A causes B. Only the last one — genuine causation — is the one most people jump to. The other three are more common than researchers like to admit.
1. Confounding Variable
A third variable C causes both A and B, creating a spurious association between A and B. Controlling for C eliminates the correlation.
2. Reverse Causation
The causal arrow points the wrong way. You assume A → B, but actually B → A. The correlation exists; the direction you assumed is wrong.
3. Spurious Correlation
No causal link, no common cause — just a coincidence that shows up in a particular dataset. Especially common when many variable pairs are tested.
4. Selection Bias
The sample was collected in a way that systematically links A and B in the data, even though they're unrelated in the full population.
The ice cream/drowning example is the most famous illustration of confounding, but reverse causation trips up researchers far more often in practice. In medical research, patients who are already sick seek more treatment, so observational studies can make treatments look ineffective or harmful when they're not. This is sometimes called "confounding by indication" — the indication for treatment is the disease severity, which also affects the outcome.
If you test 100 unrelated variable pairs for correlation at α = 0.05, you expect roughly 5 to show statistically significant results by chance alone. Spurious correlations multiply as datasets grow larger and more variables are measured. This is the statistical basis for Tyler Vigen's spurious correlations database, which documents hundreds of mathematically real but logically absurd associations.
Three Landmark Cases: What the Data Actually Showed
Case 1 — Ice Cream and Drowning: The Confounding Textbook
Ice Cream Sales and Drowning Rates
Across US counties and months, ice cream sales and swimming pool drownings are positively correlated. Both rise in June, peak in July and August, then drop sharply in September. If you ran a regression of drowning rate on ice cream sales without controlling for temperature, the coefficient on ice cream sales would be positive and statistically significant.
The confounder is temperature, and it operates through two separate pathways: heat increases outdoor swimming activity (and therefore drowning exposure) while simultaneously increasing demand for cold food and beverages. Remove the seasonal variation — compare July to July across years, or control for temperature directly — and the ice cream/drowning association disappears.
This example works as a teaching tool because the confound is obvious. In real research, confounders are rarely this transparent.
Case 2 — Smoking and Lung Cancer: From Correlation to Causation
The British Doctors Study (1951–1994)
In 1950, Richard Doll and Austin Bradford Hill published an initial observational study showing that smokers developed lung cancer at higher rates than non-smokers. The correlation was clear. Critics argued it could be confounding — perhaps people with genetic predispositions to lung cancer were also more likely to smoke, rather than smoking itself causing the cancer.
To address this, Doll and Hill began the British Doctors Study in 1951 — a prospective cohort study following 34,439 male British doctors. Over the next 40 years, they tracked who smoked, how much, and what they died of. The results left little room for alternative interpretations:
- Smokers had 14 times the lung cancer mortality rate of non-smokers
- The risk scaled with cigarette consumption — heavier smokers died at higher rates (dose-response relationship)
- Doctors who quit smoking had declining risk over time (reversibility)
- The association was consistent across different age groups, regions, and follow-up periods
This accumulation of evidence — strength, consistency, dose-response, reversibility, and biological plausibility — moved the scientific consensus from "correlation" to "causation." By 1964, the US Surgeon General's report concluded that smoking is a cause of lung cancer in men. The CDC now documents that smoking causes about 80–90% of lung cancer deaths in the United States.
Case 3 — Screen Time and Teen Depression: Still Contested
Social Media, Screen Time, and Adolescent Mental Health
A 2018 paper by Jean Twenge and colleagues used US national survey data to report that teenagers who spent more time on social media reported higher rates of depressive symptoms and lower life satisfaction. The correlation was statistically significant and attracted enormous media attention.
Subsequent analyses complicated the picture. Amy Orben and Andrew Przybylski re-examined the same datasets in 2019 using specification curve analysis — running thousands of different reasonable analytic choices — and found that social media use explained less than 0.35% of variance in adolescent wellbeing. The correlation existed but was smaller than the correlation between wellbeing and wearing glasses, eating potatoes, or sleeping with shoes on.
The reverse causation problem is also live: adolescents experiencing depression may increase social media use because they're withdrawing from in-person activities, not the other way around. Several randomized experiments where participants were assigned to reduce social media use have shown modest positive effects on wellbeing, though effect sizes are small and results are mixed.
As of 2026, this remains an active research area. The correlation is real. Whether and how much social media use causes mental health changes — versus responding to them — is not fully settled.
How Scientists Establish Causation: The Bradford Hill Criteria
Austin Bradford Hill proposed nine considerations in his 1965 presidential address to the Royal Society of Medicine. They are not a checklist where meeting all nine confirms causation, nor does failing one rule it out. They are a framework for weighing evidence. The smoking-lung cancer case satisfied most of them convincingly. That accumulation of evidence — not any single criterion — is what shifted scientific consensus.
Strength
How large is the association? Stronger associations are harder to explain away as confounding. A relative risk of 14 (smokers vs non-smokers for lung cancer) is harder to attribute to an unknown confounder than a relative risk of 1.2.
Consistency
Does the association replicate in different populations, by different investigators, in different time periods and settings? Consistent findings are harder to explain as artifacts of a particular study.
Specificity
Is the cause associated with a specific effect? A substance that only causes one particular disease is a stronger causal signal than one associated with every disease studied — the latter looks more like a confounder.
Temporality
Does the proposed cause precede the effect? This is the only criterion Hill considered necessary. If B appears before A, A cannot cause B.
Biological Gradient
Is there a dose-response relationship? If more exposure produces more effect, that is a strong signal. Smokers with higher cigarette consumption had proportionally higher lung cancer rates — a gradient that's hard to explain as confounding.
Plausibility
Is there a biologically plausible mechanism? This depends on existing scientific knowledge. Carcinogens in tobacco smoke damaging DNA in lung cells is a known mechanism — but absence of a known mechanism doesn't disprove causation.
Coherence
Does the causal interpretation conflict with what is already known about the biology and natural history of the disease? Coherence means the causal claim fits the existing scientific picture.
Experiment
Does experimental or quasi-experimental evidence support the conclusion? If removing the exposure reduces the outcome, that is stronger evidence than observation alone. Smoking cessation reducing lung cancer risk is one example.
Analogy
Are there similar cause-effect relationships already established? If other drugs in the same class cause similar outcomes, that increases plausibility. This is the weakest criterion — analogies can mislead.
Judea Pearl's Causal Hierarchy
Statistician Judea Pearl developed a mathematical framework for causal reasoning, published in his 2009 book Causality. He described a three-level hierarchy, which he calls the "causal ladder" or "ladder of causation":
Association (Seeing)
What patterns do I observe in the data? This is purely statistical correlation. "People who take aspirin more often have fewer heart attacks" is an associational statement. It makes no claim about intervention. Standard correlation analysis and regression operate at this level.
Intervention (Doing)
What would happen if I changed something? "If we give aspirin to patients, will their heart attack rate decrease?" This requires a randomized experiment or Pearl's "do-calculus" — a formal system for computing intervention effects from observational data when certain structural assumptions hold. RCTs operate at this level.
Counterfactual (Imagining)
What would have happened if things had been different? "Would this patient have had a heart attack had they not taken aspirin?" Counterfactual reasoning requires a causal model. It's the language of individual-level causation and is at the core of legal liability, clinical decision-making, and policy evaluation.
Most statistical education — including much of what's taught in introductory courses — operates at level 1. Understanding causation requires climbing to levels 2 and 3. Pearl's framework, described in more detail in his accessible book The Book of Why (co-authored with Dana Mackenzie, 2018), provides the mathematical tools to do that.
The CAUSE Framework: A Practical Checklist
The following is an original framework designed for anyone evaluating whether an observed correlation should be interpreted causally. Work through each step before drawing a causal conclusion from observational data.
Check for Correlation
Does a statistical relationship actually exist? Compute a correlation coefficient. Check for statistical significance using a p-value. Look at a scatter plot. If r is not significantly different from 0, there's nothing to explain causally.
Assess Alternative Explanations
List every plausible confounder, every reason reverse causation might apply, and every selection mechanism that could produce this correlation without a direct link. The goal is to falsify the causal interpretation before accepting it.
Understand Confounders
Identify specific variables that could jointly cause A and B. Check whether the correlation persists after statistically controlling for those variables. If controlling for a third variable reduces the A-B correlation substantially, confounding is likely.
Seek Experimental Evidence
Is there a randomized controlled trial, natural experiment, or causal inference approach (instrumental variables, difference-in-differences, regression discontinuity) that tests intervention? If A is experimentally changed and B changes accordingly, that's much stronger evidence.
Evaluate Causal Conclusions
How many Bradford Hill criteria are satisfied? Is there biological or theoretical plausibility? Does the effect replicate across different populations, datasets, and research groups? Only after working through steps C through S should you offer a causal interpretation — and even then, state it as a provisional conclusion with clear caveats.
Pearson Correlation Calculator
Enter two equal-length datasets below, separated by commas or spaces. The calculator returns the Pearson correlation coefficient r, r², the t-statistic for significance, and a scatter plot. For interpreting what r values mean, see the interpretation guide in the Pearson correlation reference page or the table in the Pearson correlation table.
Pearson Correlation Coefficient Calculator
Enter matching data points for two variables. Minimum 3 pairs required. Values must be numerical.
Scatter plot with Pearson regression line
r = [n·Σ(xy) − (Σx)(Σy)] / √{ [n·Σx² − (Σx)²] · [n·Σy² − (Σy)²] } | where n is the number of data pairs. For ranked data, use Spearman's rank correlation instead.
Interpreting the Correlation Coefficient
| r Value | Interpretation | What This Means |
|---|---|---|
| 0.90 to 1.00 | Very strong positive | Variables track each other closely in the same direction |
| 0.70 to 0.89 | Strong positive | Consistent positive relationship, considerable predictive value |
| 0.50 to 0.69 | Moderate positive | Noticeable trend, but substantial scatter around the line |
| 0.30 to 0.49 | Weak positive | Small consistent trend, many exceptions |
| 0.00 to 0.29 | Very weak / negligible | Little to no systematic linear relationship |
| −0.29 to 0.00 | Very weak negative | Small tendency to move in opposite directions |
| −0.49 to −0.30 | Weak negative | Consistent negative trend, many exceptions |
| −0.69 to −0.50 | Moderate negative | Noticeable inverse relationship with scatter |
| −0.89 to −0.70 | Strong negative | Consistent inverse relationship |
| −1.00 to −0.90 | Very strong negative | Variables track each other closely in opposite directions |
Remember that r measures linear relationship only. Variables can have a strong non-linear relationship and still show r near 0. Always plot your data before relying on r alone. The scatter plot guide covers how to read these plots in practice.
Correlation vs Causation in Practice: Four Fields
The correlation-causation mistake appears in every discipline that analyzes data. Here's how it shows up in specific fields, and what each field does (or should do) about it.
Medicine and Epidemiology
Observational studies can only establish correlation. RCTs are the gold standard for causation, but randomizing patients to harmful exposures is unethical. Epidemiologists use the Bradford Hill criteria to assess causation from observational data, and methods like propensity score matching to reduce confounding.
Economics
Economic data is almost never from controlled experiments. Economists use natural experiments, instrumental variables, and difference-in-differences to extract causal effects from observational data. For example, Card and Krueger (1994) used a minimum wage increase in New Jersey (with Pennsylvania as a control) to estimate the causal effect on employment.
Machine Learning
Most ML models are correlation machines — they learn statistical associations and make predictions based on them. A model might learn that high zip code income correlates with loan repayment and use that for decisions, without understanding the mechanism. Causal ML is an active research area addressing this gap.
Science Journalism
Headlines frequently misrepresent correlational studies as causal findings. "Coffee drinkers live longer" (a correlation) becomes "Coffee extends life" (a causal claim). Readers and journalists benefit from checking whether the cited study was observational or experimental before accepting causal language.
Before accepting a causal claim: (1) Was it an RCT? (2) Were known confounders controlled for? (3) Is there a plausible mechanism? (4) Has it replicated? If you can't answer yes to at least two of these, the study establishes correlation, not causation.
Decision Tree: Correlation or Causation?
🌳 Is This Relationship Correlation or Causation?
Glossary of Key Terms
| Term | Definition | Example |
|---|---|---|
| Correlation | A statistical relationship where two variables tend to change together | Height and shoe size are positively correlated |
| Causation | One variable directly produces change in another through a mechanism | Smoking causes lung cancer |
| Pearson r | Coefficient measuring linear correlation, −1 to +1 | r = 0.85 = strong positive linear relationship |
| Spearman ρ | Rank-based correlation coefficient for non-linear/ordinal data | Used when data isn't normally distributed |
| Confounding variable | A third variable that drives both the exposure and outcome, creating a spurious link | Heat drives both ice cream sales and drowning |
| Lurking variable | Synonym for confounder — a variable not measured or included in the analysis | Socioeconomic status is a common lurking variable in health data |
| Reverse causation | The presumed cause is actually caused by the presumed effect | People with depression may increase social media use (not the reverse) |
| Spurious correlation | A statistical association with no causal or logical basis — a coincidence | Nicolas Cage films vs. swimming pool drownings |
| RCT | Randomized controlled trial — participants randomly assigned to conditions | Drug trials where patients are randomly assigned treatment or placebo |
| Bradford Hill criteria | Nine considerations for assessing causation from observational data | Strength, consistency, dose-response, temporality, plausibility, etc. |
| Causal inference | Statistical methods for estimating causal effects from observational data | Instrumental variables, difference-in-differences, propensity score matching |
| Selection bias | Non-random sample selection creates a false association in the data | Berkson's bias in hospital-based studies |
| Dose-response | Higher exposure produces greater effect — a strong causal signal | More cigarettes smoked → higher lung cancer mortality |
| r² | The proportion of variance in Y explained by X (coefficient of determination) | r = 0.7 → r² = 0.49, meaning 49% shared variance |
| Effect size | The magnitude of an association or difference, independent of sample size | Cohen's d, odds ratio, relative risk. See the effect size guide. |
| Statistical significance | A result unlikely to occur by chance alone at a given significance level α | p < 0.05 means the result is significant at the 5% level. See significance level. |
| Observational study | Data collected without intervention — the researcher observes only | Surveying people on diet and health outcomes |
| Experimental study | Researcher actively manipulates the independent variable | Assigning participants to diet conditions and measuring outcomes |
| Natural experiment | A real-world event that mimics random assignment without researcher control | A policy change that affects only one region, used as a treatment group |
| Instrumental variable | A variable that affects the exposure but not the outcome except through the exposure | Distance to school used to estimate effect of education on earnings |
Common Misconceptions Corrected
| Misconception | Why It's Wrong | What's Actually True |
|---|---|---|
| "A strong correlation means causation is likely" | Wrong | Correlation strength tells you nothing about causation. A correlation of 0.99 could still be entirely due to confounding. |
| "If we control for enough variables, correlation = causation" | Wrong | Over-controlling can introduce collider bias. Controlling doesn't fix unmeasured confounders. |
| "RCTs are always the right method" | Overstated | RCTs have their own limitations: external validity, ethics, cost, and design. Well-designed observational studies with strong causal reasoning can be more informative in some contexts. |
| "Correlation of zero means no relationship" | Wrong | Pearson r = 0 only rules out a linear relationship. A perfectly quadratic relationship (Y = X²) can show r = 0. |
| "Large sample sizes establish causation" | Wrong | Larger samples make correlations more precise, not more causal. A study of 1,000,000 people can precisely estimate a correlation that is entirely spurious. |
| "If a study is peer-reviewed, it proves causation" | Wrong | Peer review evaluates methodology and presentation, not whether causal claims are justified. Many published observational studies make causal claims that their design cannot support. |
| "Statistical significance means practical importance" | Wrong | Significance depends on sample size. A tiny effect can be statistically significant with a large enough sample. Effect size matters more for practical conclusions. See the effect size page. |
Quick Reference: When to Say What
- Use "associated with" or "correlated with": When reporting observational data with no causal analysis
- Use "predicts" or "is linked to": For regression-based findings without experimental manipulation
- Use "may contribute to" or "is a risk factor for": When a causal interpretation is supported by multiple lines of evidence but not definitively proven
- Use "causes" or "leads to": Only when there's experimental evidence or a very strong causal case with multiple criteria met
- Avoid "X shows that Y": Unless the source material itself makes that directional claim with appropriate evidence
- Avoid "X increases your risk of Y" from a single observational study: That phrasing implies causation. Reserve it for established relationships.
Frequently Asked Questions
r = 1.0) means the two variables have a perfect positive linear relationship in the observed data, but it does not establish a cause-and-effect relationship. Two variables can be perfectly correlated because they are both driven by the same underlying factor rather than directly influencing each other.
r = [nΣ(xy) − (Σx)(Σy)] / √{[nΣx² − (Σx)²][nΣy² − (Σy)²]}. The resulting value ranges from −1 to +1, where values near ±1 indicate stronger linear relationships and values near 0 indicate little or no linear association.
Sources and Further Reading
- Hill, A.B. (1965) — "The environment and disease: Association or causation?" Proceedings of the Royal Society of Medicine, 58(5), 295–300. The original Bradford Hill paper. PMC1898525
- Doll, R. & Hill, A.B. (1954) — "The mortality of doctors in relation to their smoking habits." British Medical Journal, 1(4877), 1451–1455. The British Doctors Study foundational paper. bmj.com
- Pearl, J. (2009) — Causality: Models, Reasoning and Inference (2nd ed.). Cambridge University Press. The mathematical foundation for causal inference and the do-calculus.
- Pearl, J. & Mackenzie, D. (2018) — The Book of Why: The New Science of Cause and Effect. Basic Books. Accessible introduction to causal reasoning and the ladder of causation.
- Orben, A. & Przybylski, A.K. (2019) — "The association between adolescent well-being and digital technology use." Nature Human Behaviour, 3, 173–182. The specification curve analysis of screen time. nature.com
- Rothman, K.J. & Greenland, S. (1998) — Modern Epidemiology (2nd ed.). Lippincott-Raven. Standard reference for epidemiological causal reasoning and confounding.
- Vigen, T. — Spurious Correlations database. Hundreds of real statistical associations with no logical basis. tylervigen.com
- Stanford Encyclopedia of Philosophy — "Probabilistic Causation." Philosophical analysis of the relationship between probability and causation. plato.stanford.edu
- CDC Tobacco Data — Tobacco-related mortality statistics. cdc.gov
- American Statistical Association — Statement on p-values: Context, Process, and Purpose. amstat.org