What Pearson correlation actually measures
The Pearson correlation coefficient quantifies the strength and direction of a linear relationship between two variables, ranging from -1 (perfect inverse linear relationship) to +1 (perfect direct linear relationship), with 0 indicating no linear relationship. That word "linear" is doing a lot of work — it's the single most important qualifier in the entire concept, and it's exactly what gets dropped when people casually say "these two things are correlated."
Why a near-zero correlation doesn't mean "unrelated"
Two variables can have a strong, obvious relationship that Pearson correlation completely misses — a classic example is a U-shaped or parabolic relationship, where one variable first decreases then increases as the other grows. Despite a very clear, strong pattern, the linear correlation coefficient can come out close to zero, because a straight line fit to a U-shape captures almost none of the actual relationship. This is exactly why a scatter plot matters alongside the correlation number — visualizing the actual relationship catches patterns the single coefficient hides.
The three ways strong correlation shows up without causation
- Reverse causation — A causes B is assumed, when actually B causes A. Ice cream sales correlate with drowning incidents, but ice cream doesn't cause drowning — both increase in warm weather, which is the actual driver of both.
- Confounding variable — both A and B are driven by some third factor neither directly causes. The ice cream/drowning example above is really this: warm weather is a confounder driving both.
- Pure coincidence — especially in large datasets scanned for many possible variable pairs, some pairs will show a strong correlation purely by chance, with no meaningful relationship of any kind. This is a well-documented statistical phenomenon precisely because testing enough variable pairs makes some spurious strong correlations essentially guaranteed to appear.
Why "correlation implies causation" is such a persistent trap
Correlation is easy and cheap to compute across every pair of variables in a dataset — a correlation matrix does this automatically for an entire table at once. Causation requires a fundamentally different kind of evidence: a controlled experiment (randomly assigning which group receives a treatment), or careful causal-inference methods when experiments aren't possible. Because correlation is so much easier to obtain, it's tempting to treat a strong correlation as if it already answered the causal question — but it never does, on its own.
What would actually establish causation
Establishing that A causes B generally requires either a randomized controlled experiment (removing confounders by design, since random assignment breaks any systematic link to a third factor) or, when experiments aren't feasible, careful statistical techniques specifically designed to account for confounders and reverse-causation possibilities — techniques well beyond what a correlation coefficient alone can provide.
Common mistakes
- Treating a strong correlation as proof of a causal relationship. It's consistent with causation, but equally consistent with reverse causation, a confounding variable, or coincidence.
- Assuming a near-zero correlation means no relationship exists. It only rules out a linear relationship — a scatter plot can reveal a real, strong non-linear pattern that correlation misses entirely.
- Scanning many variable pairs and treating any strong correlation found as meaningful, without accounting for the fact that testing enough pairs makes some spurious strong correlations statistically likely to appear by chance alone.
FAQ
Does a high correlation coefficient prove one variable causes the other?
No — a strong correlation is consistent with causation but equally consistent with reverse causation, a shared confounding factor, or pure coincidence; correlation alone never establishes which is actually happening.
Can two variables have a strong relationship with a near-zero Pearson correlation?
Yes — Pearson correlation only captures linear relationships; a strong non-linear pattern (like a U-shape) can produce a correlation near zero despite an obvious, real relationship existing.
What would actually prove causation between two variables?
Generally a randomized controlled experiment, or (when that's not feasible) careful statistical techniques designed specifically to account for confounders and reverse causation — well beyond what a correlation coefficient alone provides.
Compute the full pairwise correlation matrix for your dataset with the Correlation Matrix Calculator — entirely client-side.