Ice cream sales and drowning rates are strongly correlated. Does ice cream cause drowning? Obviously not—both are driven by hot weather. This example is silly, but the underlying confusion—mistaking correlation for causation—is responsible for genuinely costly errors in medicine, policy, economics, and everyday reasoning. The distinction between statistical association and causal relationship is one of the most important and most frequently violated principles in empirical science.
What Correlation Measures
Pearson's correlation coefficient r measures the strength of linear association between two variables, ranging from −1 (perfect negative linear relationship) to +1 (perfect positive). r = 0 means no linear association—though there may be strong nonlinear relationships. Correlation is symmetric: if A correlates with B, B correlates equally with A. It is also scale-invariant: correlating height in inches vs. centimeters gives the same r. These mathematical properties make correlation a useful descriptive tool while also explaining some of its limitations as evidence for causation.
Why Correlation Doesn't Imply Causation
Three causal structures can produce correlation without direct causation. Confounding: a third variable Z causes both X and Y (hot weather causes both ice cream sales and swimming, hence both drowning and ice cream are elevated). Reverse causation: Y actually causes X rather than vice versa (sick people take medicine, but observing sick people take more medicine doesn't mean medicine causes illness). Selection bias: the sample is not representative—observing that wealthy people are healthier doesn't mean wealth causes health if both result from shared background factors.
Establishing Causation
The gold standard for establishing causation is the randomized controlled trial (RCT). Randomly assign subjects to treatment or control groups. Because assignment is random, the groups are statistically identical in all respects—including unmeasured confounders. Any subsequent difference in outcomes can only be due to the treatment. The randomization breaks all confounding paths because treatment is independent of all other variables by design. This is why RCTs are the standard for drug approval and why observational studies, however large, cannot alone establish causation.
Causal Inference Methods
When RCTs are impossible—you cannot randomize people to smoke, or countries to have different tax policies—causal inference methods attempt to recover causal effects from observational data. Instrumental variables use a variable that affects treatment but not outcome directly, breaking confounding. Regression discontinuity exploits sharp eligibility thresholds (students just above/below a test cutoff) where treatment assignment is as-good-as-random. Difference-in-differences compares changes over time between treated and untreated groups. These methods require strong assumptions that must be argued, not just assumed.
Why It Matters So Much
Confusing correlation and causation leads to real harm. Medical treatments found effective in observational studies often fail in RCTs because the association was confounded. Policy interventions based on spurious correlations waste resources or cause active harm. Financial models that confuse historical correlation with causal structure collapse when market conditions change. The replication crisis in psychology is partly driven by researchers who found correlations and published them as causal findings without adequately considering alternatives. Causal thinking is a discipline that must be actively cultivated.
Conclusion
Correlation is a starting point for scientific investigation, not its conclusion. It identifies patterns that require explanation, but the explanation might be confounding, reverse causation, or selection—not a direct causal path. The tools to move from correlation to causation—randomization, natural experiments, causal graphical models—are among the most important in the modern statistical toolkit. In an era of big data where correlations are cheap to compute and easy to mistake for insights, the discipline to ask 'but does it cause?' is more valuable than ever.