A man dances every morning at 8:01. The sun rises at 8:05. Perfect record — every dance, a sunrise. He concludes the dance causes the dawn.
You laugh, but notice what he has: a flawless correlation. Two things moving together, every single time. Correlation is exactly that — a pattern, two variables that change together in a consistent direction. What it is not, ever, is a mechanism — an account of why. The sun dancer's mistake isn't seeing the pattern. The pattern is real. His mistake is promoting a pattern to a cause.
Here are three claims that feel much more respectable than the sun dance. Each one comes from the same error.
A city installs streetlights and crime drops 42%. Lights deter crime, right? Maybe — visibility reduces opportunity. But cities installing streetlights are usually also investing in policing, community programs, economic development. Did the lights cause the drop, or were they one item in a package?
Patients who got the experimental drug had better outcomes. The drug works, right? Maybe — or doctors gave the drug to patients they judged healthy enough to handle it, in which case the drug group was healthier before the first dose. The correlation might reflect who was selected, not what was administered.
Students who receive tutoring score lower than students who don't. Tutoring harms performance? Obviously not — struggling students seek tutoring. Low scores cause tutoring-seeking. The arrow is real; it points the other way.
In every case the correlation is genuine. Whether it's causal is the entire question. So what would it actually take?
X and Y must move together. If X causes Y, they have to be correlated — no correlation, no causation. But this criterion only runs one direction: all causation requires correlation; not all correlation is causation. Covariation is necessary but not sufficient. It's the entry fee, not the proof.
The cause must come before the effect. Sounds trivial — of course causes come first. In practice it's slippery. More education leads to higher income, obviously: schooling comes before the paycheck. But family wealth precedes both — it buys the education and supplies the job connections. The time-ordering you can see may not be the time-ordering that matters.
No third variable can explain both X and Y. This is the criterion the streetlight example fails — civic investment may drive both the lights and the crime drop. It's the hardest criterion to satisfy, hard enough that it gets its own lesson. For now, hold the question: what else could be moving both of these?
Even with all three criteria in view, the arrow itself can fool you in three escalating ways.
Reverse causation — one arrow, wrong direction. People who take more medications have worse health, so medications harm health? No: sicker people take more medications. Countries with more doctors have more disease, so doctors cause disease? No: disease burden drives medical investment. Same pattern as the tutoring example — the arrow exists, you've drawn it backwards.
Bi-directional causality — two arrows, both real. Married people report being happier than single people, so marriage increases happiness? Partly — companionship and stability do help. But happier people are also more likely to marry. Neither variable is "the" cause; they cause each other. Same with poverty and poor health: poverty blocks care, food, housing; illness blocks work and piles up bills. Two true arrows at once.
Feedback loops — bi-directional causality, plus time. Anxiety keeps you awake; sleep loss dysregulates your stress response; week one's bad night becomes week four's clinical problem. The structure iterates and amplifies. This is why "just stop using social media" is hard advice: the anxiety the scrolling worsens is the same anxiety driving you back to scroll. You can't fix one variable when the other immediately re-causes it.
So the working checklist, in order: Are they correlated at all? Does the cause come first? What third thing could move both? And which way does the arrow run — one direction, both, or looping? One last thing worth noticing: every A/B test result and every "Model A beats Model B" headline is a causal claim. The criteria don't care what industry you're in.
NEXT IN THE SERIES — Levels of Measurement · back to the Library