Data Analytics: Foundations to Practice · Correlation, Relationships, and Causation
Why Correlation Doesn't Establish Causation
This is one of the most repeated phrases in data analytics, and also one of the most frequently violated principles in actual practice, because the underlying mechanisms that break the correlation-causation link are genuinely easy to overlook in the moment.
Two variables can be correlated because a third, unmeasured variable genuinely causes changes in both of them independently, rather than either variable causing the other directly; a classic illustrative example notes that ice cream sales and drowning incidents are positively correlated, not because ice cream causes drowning, but because hot weather, the actual confounding variable, independently increases both ice cream sales and swimming activity at the same time.
- A confounding variable can independently cause changes in both correlated variables, as with ice cream sales and drownings both driven by hot weather.
- Correlation cannot distinguish the direction of causation; a relationship assumed one way could actually run in the reverse direction.
- With enough variables measured over enough time, some correlations emerge purely by coincidence, with no real underlying relationship at all.
- Establishing genuine causation generally requires a controlled experiment, deliberately manipulating one variable while holding or randomizing others, unlike observing an existing correlation alone.