Back to Learning Area

Correlation vs. Causation

Why every statistician flinches when someone says "correlates with" and means "causes".

  • Fundamentals
  • 14 min
  • Beginner

Plain Language

This preview is in Plain Language. The whole portal is also available in Plain Language. You can change the language at the top. Or use the button below. Our Plain Language texts are AI-assisted and editorially reviewed. They follow the core principles of easy-to-read language, but they are not certified Easy Read: a review panel of people with learning difficulties does not test them.

Preview: Correlation and causation

What is correlation?

Sometimes two things happen together often. This is called correlation. Correlation means: both things appear together. This does not prove a cause.

An example

On hot days, people buy more ice cream. On hot days, there are also more sunburns. Ice cream and sunburns happen together. But ice cream does not cause sunburns. Both come from hot weather.

What is causation?

Causation means: one thing causes the other. This is harder to show. Researchers perform experiments. This is how they find out: is there a real cause?

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

Definition

Correlation (Pearson r)

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.

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.

Note

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

Suppose you have 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.

Caution

Misconception: No Correlation = No Connection

Pearson r only captures linear patterns. If the relationship is shaped like an inverted U — 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

Definition

Causation

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.

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.

Note

Three Criteria for Causal Claims

  1. 1. Association: A and B must be statistically related (an association is necessary but not sufficient — and it does not have to be linear).
  2. 2. Temporal order: The cause A must come BEFORE the effect B in time. If B happens first, A cannot have caused it.
  3. 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 the number of Nobel laureates per 10 million inhabitants. Suggestive? Absolutely. Causal? No. A plausible 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.

Caution

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

Definition

Confounder

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.

The wolf-husky example is an especially clear case because the confounder (snow in the background) is visually obvious. In the well-known original experiment (Ribeiro and colleagues, 2016), the model was in fact built this way on purpose: the researchers trained it on 20 hand-picked images in which every wolf stood in snow, to test whether people would spot the flaw. In tabular data, confounders are often invisible and can only be identified through statistical methods (e.g. adjustment, matching) or domain knowledge.

Note

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, especially since two series that both rise or fall over the same period often correlate highly for that reason alone. The lesson: never trust a correlation just because r is high.

Caution

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.

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 in the years that follow.
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.

Note

Key Takeaways

  1. Correlation measures co-movement — it says nothing about cause and effect. The Pearson coefficient r only captures linear relationships.
  2. 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.
  3. 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.

Check

Quiz: Correlation vs. Causation

What does a Pearson correlation coefficient of r = 0 mean?

Select one answer

Answer Key: 1) B · 2) B · 3) C · 4) B

Check

Checkpoint: Correlation vs. Causation

Can you explain this in your own words? Tick what you've got.

0 of 3 done