Correlation vs. Causation - Statistics and Probability

Section 1

Historical Context & Motivation

Humans are pattern-seeking creatures. Long before the invention of formal statistics, people noticed relationships between events—rooster crows followed by sunrise, heavy rains followed by floods—and often assumed one caused the other. This habit served our ancestors well in many cases, but it also produced countless superstitions. The formal study of correlation and causation arose precisely because scholars realized that noticing a pattern is not the same as explaining it.

1888

Francis Galton publishes Co-relations and Their Measurement, introducing the concept of correlation as a numerical measure of how two variables move together. He measures traits like height in parents and children, noticing that taller parents tend to have taller kids—but he is careful to note that the relationship is statistical, not deterministic.

1896

Karl Pearson develops the Pearson correlation coefficient (r), giving researchers a precise formula to quantify the strength and direction of a linear association. This tool becomes the backbone of statistical analysis in biology, psychology, and economics.

1950s

Researchers notice a strong positive correlation between cigarette smoking and lung cancer. Tobacco companies argue that "correlation does not equal causation," prompting statisticians like Austin Bradford Hill to develop formal criteria for inferring causation from observational data.

1965

Bradford Hill publishes his famous nine criteria for causal inference—including strength, consistency, specificity, and temporality—providing a framework that is still used in medicine and public health today.

2000s–Present

The era of big data amplifies the problem. With millions of variables to compare, spurious correlations (like the near-perfect correlation between U.S. cheese consumption and the number of people who died tangled in bedsheets) become easy to find, reinforcing the need to distinguish association from causation.

The central question this lesson addresses is deceptively simple: If two quantities tend to increase or decrease together, does that mean one is causing the other? As you will see, the answer is almost always "not necessarily," and understanding why is essential to interpreting data responsibly.

Section 2

Core Principles & Definitions

Before you can distinguish between correlation and causation, you need rock-solid definitions of both. These two ideas are related but fundamentally different, and confusing them is one of the most common errors in reasoning about data.

Correlation

A statistical relationship in which two variables tend to move together in a predictable pattern—either both increasing (positive), or one increasing while the other decreases (negative). Correlation measures association but says nothing about why the pattern exists.

Causation

A relationship in which changes in one variable directly produce changes in another. Establishing causation requires evidence beyond a statistical pattern—typically a controlled experiment, a plausible mechanism, and the elimination of alternative explanations.

Lurking (Confounding) Variable

A hidden third variable that influences both of the variables you are studying, creating the illusion of a direct relationship between them. Lurking variables are the most common reason why correlation fails to imply causation.

Spurious Correlation

A correlation that arises purely by coincidence—or through a lurking variable—and has no meaningful connection. Examples abound: per-capita margarine consumption correlates with the divorce rate in Maine, but nobody believes margarine causes divorce.

✦

Key Takeaway

Think of correlation like noticing that every time you carry an umbrella, you also wear a raincoat. The two "go together," but carrying the umbrella does not cause you to put on the raincoat. A lurking variable—rain—causes both. In statistics, the umbrella-and-raincoat pattern is the correlation; the rain is the confounding variable; and mistakenly claiming "umbrellas cause raincoat use" is the error of confusing correlation with causation.

A useful shorthand to remember is the phrase "correlation does not imply causation." This does not mean that correlated variables are never causally related—sometimes they are. It means that correlation alone is not sufficient evidence to conclude causation. You need additional reasoning, such as controlled experiments, to close that gap.

Section 3

Visual Explanation

The scatter plot below illustrates a classic scenario. Suppose a researcher collects data on hours spent studying per week and GPA for 20 high school students. The dots cluster roughly along an upward trend, suggesting a positive correlation. A line of best fit (the linear model) captures the general direction. But does more studying cause a higher GPA, or is something else going on?

The strong positive trend (r ≈ 0.93) tells us that study hours and GPA are highly correlated. But can we conclude that studying more causes a higher GPA? Not from this graph alone. Consider possible lurking variables: students who are more motivated may both study more and earn higher grades; wealthier families may provide tutors and a quiet study environment; or conscientious personality traits could drive both behaviors. The scatter plot reveals the association; it cannot, by itself, untangle which variable—if any—is doing the causing.

Section 4

Mathematical Framework

In a linear-models context, the key numerical tool for measuring correlation is the Pearson correlation coefficient, denoted r. This single number captures both the direction and the strength of the linear relationship between two quantitative variables.

Pearson Correlation Coefficient

r = Σ[(xᵢ − x̄)(yᵢ − ȳ)] / √[Σ(xᵢ − x̄)² · Σ(yᵢ − ȳ)²]

xᵢ, yᵢ = individual data values | x̄, ȳ = means of x and y | −1 ≤ r ≤ 1

Here is what r tells you: when r = 1, the data points fall perfectly on a line with a positive slope—as x increases, y increases at a constant rate. When r = −1, the points fall perfectly on a line with a negative slope. When r = 0, there is no linear association at all (though a non-linear relationship could still exist). Most real-world data produces an r somewhere in between, like 0.6 or −0.45.

Coefficient of Determination

r² = (proportion of variance in y explained by the linear model)

If r = 0.8, then r² = 0.64, meaning 64% of the variation in y is "explained" by x.

It is critical to understand what r² does not mean. When we say a linear model "explains" 64% of the variation, we are describing a statistical fit—how tightly the data hugs the line. We are not saying that x causes 64% of the change in y. A perfect correlation (r = 1) between two variables that share a confounding cause would still produce r² = 1, even though neither variable causes the other.

Linear Regression Model

ŷ = a + bx

ŷ = predicted y | a = y-intercept | b = slope = r · (s_y / s_x)

The regression equation ŷ = a + bx lets you predict a value of y given a particular x. Again, prediction is not causation. You might build a regression model that accurately predicts ice cream sales from the number of drowning incidents, but that does not mean drownings cause ice cream sales. Both are driven by a lurking variable: hot weather.

Section 5

Classifying the Relationship

When you observe a correlation between two variables, there are exactly four possible explanations for why the correlation exists. Being able to identify which explanation applies is the heart of distinguishing correlation from causation.

When you encounter a correlation in the wild—in a news article, a research paper, or a data set you're analyzing—your first instinct should be to run through all four explanations and ask which ones are plausible. Only if you can rule out reverse causation, confounding variables, and coincidence can you begin to argue for direct causation.

Strength of Evidence: From Association to Causation

Weak: Anecdote / single correlation Strong: Randomized experiment

As the spectrum above shows, evidence for causation lies on a continuum. A single observational correlation is at the weak end. Multiple consistent studies that control for confounders provide moderate evidence. A well-designed randomized controlled experiment—where subjects are randomly assigned to treatment and control groups—sits at the strong end because randomization is the gold standard for eliminating confounding variables.

Section 6

Worked Example

Let's walk through a complete scenario that puts everything together: computing a correlation, fitting a linear model, and then carefully evaluating whether causation can be claimed.

Fire Stations and Fire Damage

1
Context

A town council notices that neighborhoods with more fire stations also tend to have more fire-related property damage. They collect data from eight neighborhoods:
2
Step 1 — Compute Summary Statistics

Calculate the means: x̄ = (1+1+2+2+3+4+5+5) / 8 = 23 / 8 = 2.875 and ȳ = (42+38+61+55+82+95+118+130) / 8 = 621 / 8 = 77.625.
3
Step 2 — Compute the Correlation Coefficient

Using the formula for r, we calculate the deviations (xᵢ − x̄) and (yᵢ − ȳ), their products, and the squared deviations. After working through all eight data points: Σ[(xᵢ − x̄)(yᵢ − ȳ)] = 162.75, Σ(xᵢ − x̄)² = 18.875, Σ(yᵢ − ȳ)² = 1486.875.

r = 162.75 / √(18.875 × 1486.875) = 162.75 / √28,064.53 = 162.75 / 167.53 ≈ 0.971
4
Step 3 — Interpret the Correlation

An r of approximately 0.97 indicates a very strong positive linear relationship. As the number of fire stations increases, so does fire damage—at least in this data set.
5
Step 4 — Fit the Linear Model

The slope is b = r × (s_y / s_x). Computing the standard deviations: s_x ≈ 1.642, s_y ≈ 14.572. So b = 0.971 × (14.572 / 1.642) ≈ 8.62. The intercept is a = ȳ − b·x̄ = 77.625 − 8.62 × 2.875 ≈ 52.84.

The regression equation is: ŷ = 52.84 + 8.62x
6
Step 5 — Ask: Does This Mean Fire Stations Cause Damage?

Absolutely not! A lurking variable explains this correlation: neighborhood size (population and building density). Larger, more densely built neighborhoods have more fires simply because there are more buildings that can catch fire. The city responds by building more fire stations in those same neighborhoods. So the confounding variable (neighborhood size) drives both the number of fire stations and the amount of fire damage. If the town council misread the correlation as causation and closed fire stations to reduce damage, the result would be more destruction—the exact opposite of their goal. This example illustrates why distinguishing correlation from causation is not just an academic exercise; it has real consequences for policy decisions.

Neighborhood	Fire Stations (x)	Annual Fire Damage ($1000s) (y)
A	1	42
B	1	38
C	2	61
D	2	55
E	3	82
F	4	95
G	5	118
H	5	130

Section 7

Strengths, Limitations & Common Pitfalls

Correlation is an extraordinarily useful tool, but it has boundaries. The table below compares what correlation can do with what it cannot do, and contrasts it with what you need for a genuine causal claim.

Aspect	Correlation (Observational Data)	Causation (Experimental Evidence)
What it tells you	Two variables move together in a pattern	Changing one variable directly produces a change in the other
Study design	Observational — you measure what exists	Experimental — you manipulate one variable and randomize
Confounding variables	May be present; difficult to eliminate	Randomization distributes confounders evenly across groups
Direction of relationship	Cannot determine which variable causes the other	The manipulated variable is the candidate cause
Strength of conclusion	"X and Y are associated"	"Changing X causes a change in Y"
Real-world ease	Easy to collect; large data sets available	Often expensive, time-consuming, or ethically impossible

One important nuance: experiments are not always feasible. You cannot randomly assign people to smoke for 30 years to test whether smoking causes cancer. In such cases, researchers build a causal argument by combining multiple lines of observational evidence—large sample sizes, dose-response relationships, consistency across populations, biological plausibility, and temporal order (the suspected cause must precede the effect). These are the Bradford Hill criteria mentioned in Section 1.

✦

Key Takeaway

Think of correlation as a detective finding fingerprints at a crime scene. Fingerprints suggest the person was there (association), but they don't prove the person committed the crime (causation). To make the case, the detective needs additional evidence: a motive, a timeline, and the elimination of other suspects. Similarly, establishing causation requires going beyond the numbers and asking how and why the relationship might exist.

Section 8

Connection to Advanced Theory

The distinction between correlation and causation is not just a cautionary tale—it leads to entire branches of modern statistics. Understanding where this topic goes next can help you appreciate why the topic matters so deeply even in advanced research.

Concept You've Learned	Where It Leads
Pearson's r and linear models	Multiple regression — controlling for several variables simultaneously to isolate individual effects
Lurking / confounding variables	Causal inference — methods like propensity score matching and instrumental variables that approximate experimental conditions using observational data
Randomized experiments as the gold standard	A/B testing — widely used in tech and business to determine whether a change (e.g., a new webpage design) actually causes an improvement in user behavior
Bradford Hill criteria	Directed acyclic graphs (DAGs) — visual causal models introduced by Judea Pearl that formalize which variables to control for and which to leave alone
Spurious correlations from data mining	Bonferroni correction and false discovery rate — statistical adjustments that account for the increased chance of finding fake patterns when testing many hypotheses

In college-level statistics courses and data science programs, you will encounter these tools in depth. The key insight that carries forward is this: data alone rarely proves causation. The statistical model provides the quantitative backbone, but human judgment—about study design, possible confounders, and real-world mechanisms—is what converts a correlation into a credible causal claim.

Section 9

Practice Problems

Question 1: Problem 1 — Conceptual

A news headline reads: "Students who eat breakfast daily have higher test scores." Does this statement describe correlation or causation? Explain your reasoning.

Show answer

This statement describes a correlation. It tells us that two variables—eating breakfast and test scores—tend to move together (students who eat breakfast also tend to have higher scores). However, it does not claim that eating breakfast directly causes the higher scores. There could be lurking variables at play: for instance, students with more stable home environments may be more likely both to eat breakfast and to perform well in school. To establish causation, you would need a controlled experiment—randomly assigning some students to eat breakfast and others to skip it, then comparing test scores while holding other factors constant.

Question 2: Problem 2 — Basic Identification

A data set shows a correlation of r = −0.85 between the number of hours a student spends on social media per day and their GPA. (a) Describe the direction and strength of this correlation. (b) Calculate r² and interpret it. (c) Can you conclude that social media use lowers GPA? Why or why not?

Show answer

(a) The correlation is strong and negative. As social media hours increase, GPA tends to decrease. (b) r² = (−0.85)² = 0.7225, or about 72.25%. This means approximately 72% of the variation in GPA can be statistically "explained" by the linear relationship with social media hours. (c) No, you cannot conclude that social media use causes a lower GPA based on correlation alone. A confounding variable—such as overall self-discipline or time-management skills—could influence both variables. Students with weaker self-regulation might both spend more time on social media and study less effectively, producing the observed pattern without social media itself being the direct cause.

Question 3: Problem 3 — Intermediate

A researcher finds that cities with more coffee shops per capita tend to have lower crime rates (r = −0.72). A journalist writes: "Opening more coffee shops could reduce crime." Identify (a) the lurking variable that most likely explains this correlation, (b) explain the causal chain through which it operates, and (c) describe what kind of study would be needed to test the journalist's claim.

Show answer

(a) The most likely lurking variable is wealth or economic prosperity of the city or neighborhood. Wealthier areas attract more specialty coffee shops (because residents can afford $5 lattes) and also tend to have lower crime rates (because poverty is a well-documented driver of crime). (b) The causal chain: Higher income → more disposable income → more coffee shops open AND Higher income → less economic desperation → less crime. Coffee shops and low crime are both effects of the same cause, not cause and effect of each other. (c) To test the journalist's claim, you would ideally run a controlled experiment: randomly select a set of comparable neighborhoods, introduce new coffee shops in half of them (the treatment group), and leave the other half unchanged (the control group). Then measure crime rates over time in both groups. Because random assignment distributes confounders (like income) roughly equally between groups, any difference in crime could be attributed to the coffee shops. In practice, a natural experiment or a difference-in-differences analysis could approximate this design.

Question 4: Problem 4 — Applied / Multi-Step

A school district surveys its 12 high schools and records the average class size (x) and the average SAT score (y). The data yield r = −0.60, x̄ = 28 students, ȳ = 1120, s_x = 5, and s_y = 80. (a) Write the equation of the least-squares regression line. (b) Predict the average SAT score for a school with an average class size of 22. (c) The superintendent proposes reducing class sizes to raise SAT scores. Evaluate this proposal by identifying at least two confounding variables that could explain the observed correlation.

Show answer

(a) Slope: b = r × (s_y / s_x) = −0.60 × (80 / 5) = −0.60 × 16 = −9.6. Intercept: a = ȳ − b·x̄ = 1120 − (−9.6)(28) = 1120 + 268.8 = 1388.8. Regression line: ŷ = 1388.8 − 9.6x. (b) For x = 22: ŷ = 1388.8 − 9.6(22) = 1388.8 − 211.2 = 1177.6. The model predicts an average SAT score of about 1178. (c) The superintendent's proposal assumes that smaller classes cause higher SAT scores, but the correlation could be driven by confounders. Two likely candidates: (1) School funding per student — wealthier districts may both afford smaller class sizes and provide better resources (tutoring, test prep, experienced teachers), which boost SAT scores independently of class size. (2) Parental education and involvement — communities with more college-educated parents may demand smaller classes and also provide home environments that support academic achievement. Without controlling for these variables (through an experiment or a well-designed observational study), reducing class size alone may not produce the expected score gains.

Question 5: Problem 5 — Critical Thinking / Synthesis

A pharmaceutical company conducts an observational study and finds that patients who take its new supplement have a 15% lower rate of heart disease compared to non-users (r = −0.38 between supplement use and heart disease incidence). The company wants to advertise: "Our supplement reduces heart disease risk." (a) Explain why the observational design is insufficient to support this claim. (b) Describe in detail how a randomized controlled trial (RCT) would be designed to test the claim. (c) Even if the RCT confirms the 15% reduction, identify one remaining limitation of the conclusion.

Show answer

(a) In an observational study, participants choose whether to take the supplement. People who voluntarily take health supplements are likely to differ from non-users in systematic ways: they may exercise more, eat better diets, have higher incomes, visit doctors more regularly, and be generally more health-conscious. These confounding variables—not the supplement itself—could easily account for the 15% lower rate of heart disease. The company's claim conflates correlation with causation. (b) A proper RCT would proceed as follows: (1) Recruit a large sample of participants who are representative of the target population; (2) Randomly assign each participant to either the treatment group (receives the supplement) or the control group (receives a placebo that looks identical); (3) Ensure the study is double-blind—neither participants nor researchers know who is in which group, to prevent bias; (4) Follow both groups for a long enough period (e.g., 5–10 years) to measure heart disease outcomes; (5) Compare the heart disease rates between groups. Because random assignment distributes confounders roughly equally between the two groups, any statistically significant difference can be attributed to the supplement. (c) One remaining limitation is generalizability (external validity). Even if the RCT confirms the effect, the study's participants may not represent the broader population—perhaps they were predominantly of one age, sex, or ethnicity. The supplement might work for the study participants but not for everyone. Additionally, long-term side effects might not appear within the study window. Good science requires replication across diverse populations before a broad causal claim is considered well-established.

Summary

Lesson Summary

Summary

Throughout this lesson, you have explored one of the most consequential ideas in statistics: the distinction between correlation and causation. Correlation, measured by the Pearson coefficient r, quantifies the strength and direction of a linear association between two variables but tells you nothing about why the relationship exists. When you observe a correlation, four explanations are possible: direct causation (X → Y), reverse causation (Y → X), the influence of a confounding (lurking) variable that drives both, or pure coincidence. The coefficient of determination r² measures how much variance the linear model captures but does not prove causation. The gold standard for establishing causation is a randomized controlled experiment, in which random assignment eliminates confounders; when experiments are impractical, the Bradford Hill criteria provide a framework for building causal arguments from observational evidence.

The practical takeaway is simple but powerful: whenever you encounter a claim that one thing causes another, ask whether the evidence is correlational or experimental. Look for lurking variables, consider reverse causation, and remember that in a world awash in data, spurious correlations are everywhere. Thinking critically about these distinctions will make you a better consumer of research, a sharper data analyst, and a more informed citizen.