Correlation Is Not Causation, With Real Examples

Two lines moving together is where the mistake starts

You've seen the graph a hundred times: two lines that rise and fall in step, and a caption implying one drives the other. Ice cream sales and drownings. Screen time and unhappiness. A country's chocolate consumption and its Nobel prizes. The lines really do move together — that part is usually true. The caption is where it goes wrong.

Correlation means two things move together in a measurable way. Causation means changing one actually changes the other. The whole trap is that correlation is easy to see in data and causation is almost impossible to see in data — so people see the first and assume the second, because a graph that moves together feels like proof.

This is the most consequential reasoning error you can learn to catch, because it's the engine behind bad health advice, bad policy, bad investing, and half of what you'll read online. The fix isn't cynicism. It's a single question you ask every time two things are linked.

Whenever A and B move together, there are exactly four explanations

That's the whole tool. If A correlates with B, it's because of one of these — and your job is to work out which:

1. A causes B. The obvious one. Turning the tap causes water to flow.
2. B causes A (reverse causation). The arrow points the other way from what the story assumes.
3. A third thing, C, causes both (a confounder). A and B don't touch each other at all; something else drives them both, so they rise and fall together.
4. Coincidence. With enough data series in the world, some will line up by pure chance and mean nothing.

A correlation on its own cannot tell you which of the four you're looking at. That's not a limitation you can compute your way around with more data — it's built into what a correlation is. So the reflex is: see a link → run the four → ask which one this is.

The four, with real examples

Confounding — the one you'll meet most

Ice cream sales correlate with drownings. Both go up in the same months. Does ice cream cause drowning? No — summer causes both. Hot weather sells ice cream and sends people into the water. Summer is the confounder (C), sitting behind both, and the ice-cream–drowning link is real, strong, and completely non-causal.

Children's shoe size correlates with reading ability. Kids with bigger feet read better. Bigger shoes don't teach reading — age does both. Older children have bigger feet and can read more.

Confounders are everywhere the moment you look, and they're the reason "controlling for" other variables is the entire job of serious data work: you're trying to hold C still and see if the A–B link survives.

Reverse causation — the arrow points the other way

Studies find that people who feel in control of their lives are healthier. The tempting story: feeling in control makes you healthy. At least as plausible: being healthy makes you feel in control. Or a study notes that struggling companies use more consultants — do consultants cause trouble, or does trouble bring in consultants? The correlation is symmetric; the arrow isn't, and the data alone won't tell you its direction.

Coincidence — the spurious ones

Tyler Vigen's collection is famous for a reason: US cheese consumption tracks the number of people who die tangled in their bedsheets; the divorce rate in Maine tracks per-capita margarine consumption. Tight, beautiful correlations, all meaningless. Run enough variables against each other and chance alone produces stunning matches. This is why a single striking correlation, with no mechanism behind it, is worth nothing.

And sometimes A really does cause B

Smoking and lung cancer correlate — and here the arrow is real. But notice how that was established: not from the correlation, but from a mechanism (what smoke does to lung tissue), a dose–response relationship (more smoking, more risk), the right time order, and the link surviving every attempt to explain it away with a confounder. Causation had to be earned, over decades. The correlation was only the starting question.

How causation actually gets established

You don't prove cause from a correlation. You prove it by ruling the other three out:

• The randomised controlled trial (RCT). Split people randomly into two groups, change one thing for one group, hold everything else equal. Randomisation is what kills confounders — it scatters every C evenly across both groups, so anything left is down to the thing you changed. This is why "gold-standard evidence" means an experiment, not a survey.
• Natural experiments. When you can't randomise, you look for a moment where something changed as if at random — a law that took effect in one region and not the next.
• Mechanism. A plausible chain of how A would cause B. No mechanism, be suspicious.
• Dose–response. More of A, more of B. Genuine causes often (not always) show this.
• Time order. The cause has to come before the effect. Obvious, and constantly ignored.

When you read "linked to", "associated with", "correlated with" — those are the honest words for we found a correlation and haven't shown cause. When you read "causes", ask which of the above they did.

What this means for you

• See a link, run the four. A→B, B→A, C→both, or chance. The correlation can't tell you which.
• Reach for the confounder first — it's the most common and the most invisible. What third thing could drive both? (Ask "what does summer explain here?")
• Check the arrow. Reverse causation is easy to miss because the wrong direction often sounds natural.
• No mechanism + no experiment = not causation, however tight the graph. A striking correlation with no how behind it may be pure coincidence.
• "Associated with" is a confession. It's the careful writer telling you they found correlation and stopped there.

Try it: take apart four causal claims (30 min)

1. Collect four claims of the form "A causes B" from real life — a headline, an advert, a wellness tip, a político or influencer, a "people who do X are more likely to Y" study.
2. For each, run the four explanations and write:

CLAIM: "A causes B" .............................................
1. Could A cause B? (mechanism?) ................................
2. Could B cause A? (reverse) ...................................
3. What C could cause both? (the confounder) ...................
4. Could it be coincidence? (any mechanism at all?) ............
MOST LIKELY EXPLANATION: ........................................
WHAT EVIDENCE WOULD SETTLE IT: (what experiment/data?) .........

3. For at least one, name the specific confounder you think is really doing the work.

✅ Finish check: four claims, each with all four explanations considered, the most likely one named, and — for at least one — the confounder identified and the experiment that would settle it.

Summary card

• Correlation = moves together. Causation = changing one changes the other. Correlation is visible in data; causation almost never is.
• Four explanations for any A–B link: A→B, B→A, C→both, coincidence. The correlation can't pick.
• Confounding is the default suspect — a third thing (summer, age, wealth) driving both.
• Reverse causation — the arrow may point the opposite way to the story.
• Coincidence — enough series guarantees spurious matches; no mechanism, no trust.
• Cause is earned by experiment (RCT), natural experiment, mechanism, dose–response and time order.
• "Associated with / linked to" = correlation only. "Causes" = show me the experiment.

Sources

1. Pearl, J. & Mackenzie, D. — The Book of Why, 2018
2. Reichenbach, H. — The Direction of Time (common-cause principle), 1956
3. Vigen, T. — Spurious Correlations, 2015
4. Fisher, R. A. — The Design of Experiments, 1935

Next lesson: DT-09 — A Four-Step Protocol for Verifying a Source (L1) Related: DT-02 How Charts Lie · DT-01 Reading Data · DT-08 Logical Fallacies · MN-10 The Anatomy of a Scam Path: Hard to Fool — 3/6