Skip to content

Correlation vs. Causation Testing

Why It Matters

The human mind is a correlation-detection machine evolved for a world where most correlations were causally meaningful. In modern environments full of coincidental correlations, spurious relationships, and common-cause confounders, this instinct systematically misleads. We see illness correlating with medicine and conclude medicine causes illness (reverse causation). We see ice cream sales correlating with drowning and conclude ice cream causes drowning (common cause: hot weather). We make policy, personal decisions, and predictions based on these errors.

The Key Move

When observing correlation between A and B, systematically test four hypotheses:

  1. A causes B,
  2. B causes A,
  3. C causes both A and B (common cause),
  4. Coincidence/selection bias/measurement artifact. For each hypothesis, ask: what would we observe if this were true that we wouldn't observe under the other hypotheses? Design or find natural experiments that distinguish them.

Classic Application

Medical studies. Hormone replacement therapy (HRT) correlated with better cardiovascular health in observational studies. Randomized trials showed HRT actually increased heart disease risk. The correlation existed because healthier, wealthier women were more likely to seek HRT (selection bias), not because HRT protected hearts. Only experimental manipulation distinguishes correlation from causation.

Surprising Application

Personal productivity. "I'm most productive when I work late at night" might reflect:

  1. Night work causes productivity,
  2. Being energized causes both staying up late and productivity,
  3. Only important work happens at night because you procrastinate during the day (reverse causation), or
  4. Memory bias - you remember dramatic late-night sessions, forget mundane day work. Testing requires tracking productivity metrics across different schedules while controlling for task difficulty and motivation.

Failure Modes

Assuming randomized controlled trials are always feasible or ethical (they often aren't). Treating correlation as worthless evidence (it's weak evidence for causation, strong evidence for further investigation). Missing bidirectional causation (A causes B AND B causes A). Confusing "not proven causal" with "proven non-causal." Over-reliance on temporal order (A before B doesn't prove A causes B - common cause C might precede both).

Go Deeper

Pearl, "The Book of Why" on causal inference; Hill, "The Environment and Disease: Association or Causation?" (Bradford Hill criteria)