Interpretation discipline

Correlation vs Causation

Distinguish association from causation by checking confounding, reverse direction, and study design limits before making causal claims from correlated data.

How this page is maintained

Written for learners, checked against the sources below, and reviewed every year. Last reviewed July 22, 2026.

Short answer

Correlation measures association, not cause and effect. A nonzero correlation can result from direct causation, reverse causation, confounding variables, or coincidence, so causal conclusions require stronger design or assumptions.

  • Association alone is not a causal proof.
  • Confounders can create or mask relationships.
  • Causal claims need design evidence such as randomization or credible identification strategy.

What correlation actually quantifies

Pearson correlation r summarizes linear association between two quantitative variables, from -1 to 1. Values near +/-1 indicate stronger linear pattern; values near 0 indicate weak linear pattern.

Correlation is sensitive to outliers and does not capture nonlinear relationships well. A scatterplot should always accompany reported r values.

Why causal interpretation is risky

A third variable can influence both X and Y, creating confounding. For example, seasonal effects can move demand and staffing together without one causing the other directly.

Causal inference typically needs experimental control, quasi-experimental design, or strong assumptions with supporting diagnostics.

  • Ask what alternative explanations fit the same association.
  • Check temporal order before causal language.
  • Use precise phrasing such as associated with when causality is not established.

Interpret association between training hours and output

A team dataset shows r=0.62 between monthly training hours and monthly output.

  1. Plot a scatter diagram and inspect for linear pattern and outliers.
  2. List plausible confounders such as role seniority and project complexity.
  3. Fit a simple model with and without a key confounder and compare coefficient changes.
  4. Report finding as association unless design supports causal identification.
Result: The observed positive association is real in the sample, but causal effect is not established from correlation alone.

Common mistakes

  • Using correlated with as if it means caused by.
  • Ignoring confounding variables.
  • Reporting r without a scatterplot.
  • Assuming r near 0 means no relationship of any form.

Try one

Can a strong correlation prove causation by itself?

No. Strong association can still come from confounding, reverse direction, or other non-causal mechanisms.

Sources

Learn this with a tutor

Tell LearnLive what you already know and what you need to do with correlation vs causation.

Build this course