Causation vs correlation
Why two numbers that move together may not drive each other, the three patterns behind every relationship (chain, fork, collider), and a checklist before you act on a correlation.
The difference in one line
Correlation: A and B move together. Causation: changing A changes B. The first is a fact about data you have seen; the second is a claim about what happens when you act.
Four reasons two numbers move together
- A causes B. More staff on shift, fewer late deliveries.
- B causes A (reverse causation). Stores with more sales hire more staff; staff numbers follow sales, not the other way round.
- Something else causes both (a common cause, or confounder). The season drives both ad spend and sales.
- Selection. You only look at rows that passed a filter that depends on both. Among hired candidates, test scores and interview scores can look unrelated or even opposed, because people weak in both were never hired.
Only the first reason means that acting on A will move B.
Three patterns behind every relationship
| Pattern | Shape | A and C before you account for B | After you account for B |
|---|---|---|---|
| Chain | A → B → C | Related | Unrelated: B carries the whole effect |
| Fork | A ← B → C | Related | Unrelated: B was the common cause |
| Collider | A → B ← C | Unrelated | Related: accounting for B creates a false link |
These patterns are what causal discovery methods look for. They are also why "control for everything" is bad advice: controlling for a collider, or for something caused by the outcome, adds bias instead of removing it.
How causation is established
- Experiments (A/B tests, randomised trials) set A by chance, so nothing else can explain the difference. They are the strongest evidence, but often slow, costly or impossible.
- Causal discovery reads observational data and proposes which links are causal, using the patterns above. It narrows down what to test.
- Domain knowledge rules out what cannot happen: sales today cannot cause last month's weather.
In practice they work together: discovery proposes, knowledge filters, and an experiment confirms the links that matter most.
Checklist before acting on a correlation
- Could the arrow point the other way?
- Is there a common cause that drives both (season, size, region, a busy day)? Is it in your data?
- Were rows filtered on something related to both?
- Does the relationship hold in different periods or segments?
- If you changed A tomorrow, by what mechanism would B change?
- Is there a cheap way to test it: a pilot, a holdout group, one region first?
Try it in UCMT Labs
Upload a table with the variables you suspect, including likely common causes, and run causal discovery. A link that disappears once the other columns are taken into account was a correlation; a link that stays, with a direction, is a candidate cause to test. See What is causal discovery? and How to read your results.