Why every statistician flinches when someone says "correlates with" and means "causes".
Fundamentals 14 min Beginner May 10, 2026
Countries that eat more chocolate win more Nobel Prizes. The more firefighters a city has, the more fires it reports. The data is real. The conclusion is wrong.
The gap between "these things move together" and "one causes the other" is the most important lesson in data literacy. This article shows you how to distinguish real connections from statistical illusions.
Correlation: Co-Movement, Not Causation
Correlation (Pearson r)
AnalogyDefinition
On days when ice cream sales are high, hospitals report more sunburn cases. The two variables correlate positively — a scatterplot would show a clear upward trend. But eating ice cream does not cause sunburn. Both are driven by a third factor: sunny, hot weather.
Analogy:
On days when ice cream sales are high, hospitals report more sunburn cases. The two variables correlate positively — a scatterplot would show a clear upward trend. But eating ice cream does not cause sunburn. Both are driven by a third factor: sunny, hot weather.
Definition:
Correlation measures how strongly two numerical variables move together. The Pearson correlation coefficient r summarises this as a single number: r = +1 means perfect positive linear correlation, r = -1 perfect negative, r = 0 no linear relationship. Important: r = 0 does NOT mean "no connection" — it means no LINEAR connection.
The ice cream/sunburn analogy works because the absurdity of "ice cream causes sunburn" is immediately obvious. But beware: in real datasets — medical studies, AI training data, financial models — the hidden third variables are often subtle and sometimes unmeasurable.
Scatterplot Patterns: What r Means
Strong positive (r ≈ +0.9) Points rise from lower-left to upper-right. When A goes up, B goes up too.
Strong negative (r ≈ -0.9) Points fall from upper-left to lower-right. When A goes up, B goes down.
No linear (r ≈ 0) Points scattered like a cloud. Changes in A tell you nothing about B (linearly).
Example: Firefighters and Fires
Data from 10 cities: number of firefighters vs. number of fires per year. A scatterplot reveals a strong positive correlation (r ≈ 0.85). Naive interpretation: "More firefighters cause more fires." Correct interpretation: larger cities have more buildings (more fire risk) AND more firefighters. The confounder is city size.
Misconception: No Correlation = No Connection
Pearson r only captures linear patterns. If the relationship is U-shaped — for example, stress vs. performance: low stress = low performance, medium stress = high performance, high stress = low performance — r can be close to 0 despite a strong, meaningful pattern.
Causation: When A Actually Causes B
Causation
AnalogyDefinition
You have a headache. You take a pill. An hour later, the headache is gone. Did the pill cause the relief? Maybe — but maybe the headache would have gone away on its own (natural recovery). Maybe just believing you took medicine helped (placebo effect). To know if the pill ACTUALLY works, you would need a clone of yourself: one takes the pill, one does not. Since cloning is impractical, randomised controlled trials (RCTs) approximate this ideal.
Analogy:
You have a headache. You take a pill. An hour later, the headache is gone. Did the pill cause the relief? Maybe — but maybe the headache would have gone away on its own (natural recovery). Maybe just believing you took medicine helped (placebo effect). To know if the pill ACTUALLY works, you would need a clone of yourself: one takes the pill, one does not. Since cloning is impractical, randomised controlled trials (RCTs) approximate this ideal.
Definition:
Causation means that changing variable A systematically produces a change in variable B — a real cause-and-effect relationship, not just a pattern in the data. Three conditions must ALL be met before a causal claim is justified.
The pill analogy simplifies several aspects: in real life, multiple factors often act simultaneously (other medications, sleep, stress), and effects can be delayed or dose-dependent. The clone idea is also an idealisation — in practice, RCTs work with large groups and statistical methods to average out individual differences.
Correlation says:
A and B move together. This is an observation — it says nothing about whether A causes B, B causes A, or a third factor drives both.
Causation says:
A change in A directly produces a change in B. This is a much stronger claim — it requires association, temporal order, AND confounder exclusion.
Three Criteria for Causal Claims
1
Association: A and B must be statistically related (correlation is necessary but not sufficient).
2
Temporal order: The cause A must come BEFORE the effect B in time. If B happens first, A cannot have caused it.
3
Confounder exclusion: No third variable C should better explain the relationship. All plausible alternative explanations must be ruled out.
Example: Chocolate and Nobel Prizes
A study by Franz Messerli in the New England Journal of Medicine (2012) found a correlation of r ≈ 0.79 between per-capita chocolate consumption and Nobel Prize count per country. Suggestive? Absolutely. Causal? No. The confounder is national wealth (GDP per capita): wealthier countries can afford more chocolate AND more research funding. Without controlling for wealth, the chocolate-Nobel correlation is meaningless for causal claims.
Misconception: High Correlation + Plausible Story = Causation
A convincing narrative does NOT replace rigorous causal analysis. Humans are pattern-matching machines — we instinctively build stories around correlations. The discipline of statistics is, in large part, the discipline of resisting premature causal stories.
Confounders: The Hidden Third Variable
Confounder
AnalogyDefinition
An image classifier is trained to distinguish wolves from huskies and achieves high accuracy. Investigation reveals: the model primarily detects SNOW in the background. Wolf photos were mostly taken in snowy wilderness; husky photos in backyards and living rooms. The model learned "snow = wolf" — a spurious correlation driven by the confounder "photo environment." Feed it a wolf on grass or a husky in snow, and it fails spectacularly.
Analogy:
An image classifier is trained to distinguish wolves from huskies and achieves high accuracy. Investigation reveals: the model primarily detects SNOW in the background. Wolf photos were mostly taken in snowy wilderness; husky photos in backyards and living rooms. The model learned "snow = wolf" — a spurious correlation driven by the confounder "photo environment." Feed it a wolf on grass or a husky in snow, and it fails spectacularly.
Definition:
A confounder is a hidden third variable C that influences BOTH the presumed cause A and the presumed effect B, creating a spurious (false) correlation between A and B. A spurious correlation is a statistical relationship that exists purely due to confounders or chance, with no plausible causal mechanism.
The wolf-husky example is an especially clear case because the confounder (snow in the background) is visually obvious. In tabular data, confounders are often invisible and can only be identified through statistical methods (e.g. adjustment, matching) or domain knowledge.
Absurd Correlations from Real Data
Tyler Vigen's website (tylervigen.com) catalogues absurd correlations from real data: the divorce rate in Maine correlates almost perfectly with per-capita margarine consumption. Pool drownings correlate with Nicolas Cage film appearances. These pairs have no causal connection — they correlate purely by chance. With thousands of time series, high correlations are statistically inevitable. The lesson: never trust a correlation just because r is high.
Misconception: AI Models Understand Causation
Standard ML models learn whatever correlations maximise training accuracy. They have no concept of cause and effect. This is why models can fail catastrophically in new environments — the correlations they relied on may not hold. Understanding this distinction is the first step toward building robust AI.
Interactive: Exposing Famous Spurious Correlations
Click the cards to uncover famous spurious correlations. Behind every convincing-sounding claim hides a confounder or pure chance.
Misconception #1
Ice cream sales cause drowning — on days with high ice cream sales, more people drown.
Correction
The confounder is hot weather. On hot days, more people buy ice cream AND go swimming. The ice cream has nothing to do with drowning.
Classic statistics textbook example (Utts & Heckard, 2015)
Misconception #2
Chocolate makes you smart — countries with higher chocolate consumption win more Nobel Prizes.
Correction
The confounder is national wealth (GDP per capita). Wealthier countries can afford more chocolate AND more research funding. Messerli (2012) published this in the NEJM deliberately as a warning against premature conclusions.
Messerli, New England Journal of Medicine (2012)
Misconception #3
Nicolas Cage films kill — the number of his film appearances correlates with swimming pool drownings.
Correction
Pure chance. With thousands of available time series, high correlations inevitably appear without any causal connection. Tyler Vigen's project shows dozens of such absurd pairs.
Storks deliver babies — Europe's stork population correlates with the birth rate.
Correction
The confounder is rurality. Rural areas have both more storks (nesting sites) and historically higher birth rates. Matthews (2000) analysed this correlation as a statistical teaching example.
Matthews, Teaching Statistics 22(2), 2000
Deep Dive: Spurious Correlations in AI
AI models optimise for any pattern in training data, whether it reflects reality or not. Features like zip code or image background can become proxy variables for confounders. When conditions change (distribution shift), these learned shortcuts break down. This is why understanding correlation vs. causation is not just academic — it is a practical prerequisite for building reliable AI systems.
Self-Test: Correlation or Causation?
Five scenarios from research and everyday life. For each one, decide: Is this causation, or just correlation?
0/5 answered
Question 1
Smoking and lung cancer: Smokers develop lung cancer significantly more often than non-smokers.
Question 2
Children with bigger shoes read better than children with smaller shoes.
Question 3
People who use sunscreen more frequently are diagnosed with skin cancer more often.
Question 4
Countries with more televisions per household have higher life expectancy.
Question 5
In a randomised, double-blind study, aspirin significantly reduces heart attack risk compared to placebo.
Key Takeaways
Correlation measures co-movement — it says nothing about cause and effect. The Pearson coefficient r only captures linear relationships.
For a causal claim, you need three things: association, temporal order, and confounder exclusion. Without all three, any correlation is just a hint, not a conclusion.
AI models are correlation machines — they learn every pattern in the data, whether causal or not. Understanding this difference is the first step toward AI literacy.
Quiz: Correlation vs. Causation
Question 1 / 4
What does a Pearson correlation coefficient of r = 0 mean?
1. What does a Pearson correlation coefficient of r = 0 mean?
☐ A) There is no connection between the variables
☐ B) There is no LINEAR connection — other patterns (e.g. U-shaped) are still possible
☐ C) The variables do not cause each other
☐ D) The data is flawed
2. A study shows: children who take music lessons have better grades. Which confounder could explain this correlation?
☐ A) Music lessons make children smarter
☐ B) Higher-income families can afford music lessons AND provide more academic support for their children
☐ C) Children with good grades ban themselves from music
☐ D) Grades and music measure the same skill
3. A company finds that teams with higher satisfaction also have higher revenue. Which of the three causation criteria is hardest to verify here?
☐ A) Association (statistical connection)
☐ B) Temporal order (what came first?)
☐ C) Confounder exclusion (is there a better explanation?)
☐ D) All three are equally difficult
4. An image recognition model identifies wolves with 95% accuracy on the test dataset. On new photos from a zoo, it fails. Why?
☐ A) The model simply needs more training data
☐ B) The model learned a spurious correlation (e.g. snowy background) instead of the animal itself
☐ C) Zoo photos have lower resolution
☐ D) 95% accuracy is not sufficient
Answer Key: 1) B · 2) B · 3) C · 4) B
Checkpoint: Correlation vs. Causation
I can explain what the Pearson correlation coefficient r measures and what its range is (-1 to +1).
I can name the three criteria that must be met for a causal claim (association, temporal order, confounder exclusion).
I can explain with an example what a confounder is and why AI models that learn correlations can fail in new environments.