Correlation and Causation: The Difference, With Examples

Correlation means two variables move together; causation means changing one makes the other change. Every correlation you find has exactly four possible explanations — real causation, reverse causation, a confounding variable, or coincidence — and the data alone usually can't tell you which one you're looking at.

Why doesn't correlation imply causation?

Because a correlation coefficient is symmetric and mechanism-blind: it measures co-movement, not direction or cause. The same r = 0.9 is produced whether X drives Y, Y drives X, hot weather drives both, or you got lucky in a small sample. Observational data shows you what happened together, never what would have happened had you intervened — and that counterfactual is what causation means.

The Four Explanations for Any Correlation

When X and Y are correlated, work through this checklist before claiming anything:

1. Real Causation (X → Y)

Sometimes the obvious story is true. Smoking correlates with lung cancer because smoking causes lung cancer — a conclusion cemented by dose-response evidence, mechanism, and experiments in animals. Real causation exists; the point is that it's one hypothesis among four, not the default.

2. Reverse Causation (Y → X)

You have the arrow backward. Towns with more hospitals have more deaths — not because hospitals kill people, but because sick people go where hospitals are. Likewise, "people who carry lighters get lung cancer more often": the smoking causes the lighter-carrying, not the reverse.

Correlation: hospitals & deaths
Wrong: Hospitals cause deaths
Right: Illness causes both hospital visits and deaths

3. Confounding Variable (Z → X and Z → Y)

A hidden third variable drives both. The classic: ice cream sales and drowning deaths correlate at r ≈ 0.95. Ice cream doesn't drown anyone — hot weather causes both more ice cream and more swimming. Confounding is the most common trap in business data, and it can even flip a trend's direction entirely, as in Simpson's paradox.

Hot weather → More ice cream sales
Hot weather → More swimming → More drownings

Ice cream and drownings correlate — with zero causation between them

4. Pure Coincidence

With enough variable pairs, some correlate by chance. Nicolas Cage film releases correlate with pool drownings (r = 0.67); US cheese consumption tracks deaths by bedsheet entanglement. Test 1,000 random pairs at p < 0.05 and about 50 will "correlate." Small samples and cherry-picked time windows make spurious hits far more likely.

How do you prove causation?

The gold standard is a randomized controlled experiment: randomly assign units to treatment and control, so confounders balance out on average and any outcome difference must come from the treatment. When you can't randomize, causal inference uses time order, dose-response, mechanisms, and statistical controls — evidence that strengthens a causal claim without fully proving it.

1. Controlled Experiment (Gold Standard)

• Group A: gets the treatment (new feature)
• Group B: control (no change)
• Random assignment breaks every confounder → a difference is causal evidence

2. Time Order

The cause must come before the effect:

• ✅ Marketing spend (Monday) → Sales increase (Tuesday)
• ❌ Sales (Monday) → Marketing spend (Tuesday) can't be causal

3. Dose-Response Relationship

1 hour study → +5 points
2 hours study → +10 points
3 hours study → +15 points

Consistent dose-response strengthens the causal claim

4. Plausible Mechanism

• ✅ Exercise → Heart health (clear biological mechanism)
• ❌ Shoe size → Reading ability (no mechanism; age confounds both)

5. Eliminate Confounders Statistically

# Regression controlling for known confounders
model = ols('sales ~ marketing + temperature + day_of_week + holidays')

# Significant marketing coefficient after controls → stronger causal claim
# (but only for confounders you measured!)

Controls only handle confounders you thought of and measured — the deeper reasons this is hard are covered in why causal inference is hard.

Practical Questions to Ask

Question	What It Checks
Does X happen before Y?	Time order
What else changed at the same time?	Confounders
Could Y cause X instead?	Reverse causation
How many pairs did we test to find this?	Coincidence / multiple comparisons
Can we run an experiment?	True causation test
Does the relationship make sense?	Plausibility

Common Business Examples

Email Open Rates and Revenue

Correlation: High email opens → High revenue

Confound: Engaged customers do both

Test: A/B test email content to prove causation — see A/B testing explained

Website Visitors and Sales

Correlation: More traffic → More sales

Could be reverse: More sales → More word of mouth → More traffic

Test: Run paid ads (controlled traffic increase)