Correlation vs Causation: Moving Together Isn't Caused By

What you'll learn

Explain the difference between "moves together" and "causes" without hand-waving
Name the three usual suspects: confounders, reverse causation, and selection effects
Know what it actually takes to demonstrate causation — and when to say "let's test it"

“Customers who use the analytics feature churn 40% less. Let’s push every customer into the analytics feature!” It’s the most natural sentence in business — and it might be exactly backwards. Maybe analytics keeps customers around. Or maybe the customers who were already committed are the ones who bother setting up analytics, and pushing indifferent customers into it will achieve nothing except annoying them. The data is identical either way. This module is about the gap between those two worlds — the most expensive gap in data literacy.

The words that matter

Correlation: Two things tend to move together — when one is high, the other tends to be high (or low). It’s a pattern, not an explanation.
Causation: Changing one thing actually changes the other. This is what your decision needs — and what correlation alone can’t prove.
Confounder: A hidden third factor driving both things at once, creating the illusion that they drive each other.
Reverse causation: The arrow points the other way — Y causes X, not X causes Y.
Selection effect: The groups you’re comparing were different before the thing you’re studying — the users who chose a feature aren’t like the users who didn’t.
Survivorship bias: Only the winners are still around to be measured, so their shared habits look like the recipe for winning.

The three boring explanations

Ice cream sales and drowning deaths rise and fall together, every single year. Ice cream does not cause drowning. Summer causes both — more swimming, more dessert. That hidden third factor is a confounder, and the corporate version is everywhere: teams with more documentation ship faster (or do well-run teams both document and ship?); employees who take training get promoted more (or do ambitious people both sign up for training and chase promotions?).

Reverse causation flips the arrow. “Companies with big data teams grow faster — hire data scientists!” Perhaps. Or perhaps growing fast generates the money and the data problems that lead companies to hire data teams. Same correlation, opposite advice.

Selection effects and survivorship round out the trio. Beta users love the new interface — but beta users are your enthusiasts; their opinion tells you little about the indifferent majority. And every airport bookstore hit listing “the seven habits of billion-dollar companies” is built on survivorship bias: nobody interviews the dead companies that had the same seven habits. The World War II version is the famous one — armor the planes where returning bombers weren’t hit, because the ones hit there never returned.

Two lines rise together and the brain writes a story. Then the confounder fades in — driving both, connecting neither.

Text description of this diagram

Two lines climb together across the chart — blue feature usage and green retention — and a red “causes?” flashes between them, the story every dashboard viewer instinctively writes. Then the real explanation fades in above: an amber bubble labeled the hidden third thing: customer commitment, with dashed arrows reaching down to both lines — “drives this… and this.” Committed customers both adopt features and stick around. The two lines never touch each other at all; they’re both symptoms of the thing in the bubble.

What it takes to actually prove it

Here’s the constructive part. Correlation plus a plausible story still isn’t causation — but causation can be demonstrated. The gold standard: intervene and compare. Take a group of similar customers, push half of them (chosen at random) into the analytics feature, leave the other half alone, and watch what happens to churn in each group. Randomness is the magic ingredient — it makes the two groups alike in every other way, so the only systematic difference is the thing you changed. If churn drops in the pushed group, you’ve earned the word “causes.”

That procedure has a name you’ll meet properly soon — an A/B test — and it gets the full plain-English treatment in module 9. For now, the important skill is cheaper: when someone leaps from “correlates” to “therefore we should,” you don’t need to win the argument. You just need the sentence: “Interesting pattern — before we bet the roadmap on it, can we test whether it’s causal?” That sentence has saved companies millions, and it’s free.

Common misunderstanding

“The data speaks for itself.” Data never speaks for itself — it answers exactly the question it was asked, from exactly the group it was collected on. “Feature users churn less” is the data. “The feature reduces churn” is an interpretation — one of at least three, and the flattering one gets chosen suspiciously often. When the interpretation conveniently supports the roadmap of the person presenting it, that’s precisely when the boring explanations deserve a hearing.

Sort it: what’s really going on?

Each claim below jumps from a real correlation to a conclusion. Sort each one by the most likely boring explanation — or decide it’s worth testing as causal.

Confoundera third thing drives both

Reverse causationthe arrow points backwards

Survivorshiponly winners got measured

Tip: drag with a mouse, or tap an item then tap a bucket on touch screens. Get one wrong and the answer key appears.

Try this at work

When a correlation is presented as a reason to act, run the three-suspects check out loud: Could a third factor drive both? Could the arrow point backwards? Are we only looking at survivors? If the claim survives all three, it’s promoted from “story” to “hypothesis” — and hypotheses earn tests, not roadmaps.

The bottom line

“Moves together” is a pattern; “causes” is a claim about what happens if you intervene. Between them stand three boring explanations — confounders, reverse causation, survivorship — and the only clean way past all three is a controlled test.

Why it matters

Nearly every “the data shows we should…” argument you’ll hear is a correlation wearing causation’s jacket.

Roadmaps, budgets, and reorgs get justified this way weekly — and the AI era raises the stakes, because machine-learning models are correlation engines by construction (the AI & Automation course shows what that means in practice). People who can say “confounder?” at the right moment aren’t being negative; they’re the reason the company tests before it bets. Which brings us to the machinery of testing — but first, one more source of shaky numbers: surveys, and the art of asking who was actually asked.

Quick check

1. Ice cream sales and drownings rise together every year because…

2. "Feature users churn less, so the feature reduces churn." What's the selection-effect objection?

3. What's the cleanest way to find out whether the feature actually reduces churn?

Answers explained

C is correct — hot weather sends people swimming and buying ice cream; the two symptoms track each other without touching. (If you picked A: no lifeguard has ever blamed dessert. If you picked B: it’s too regular for coincidence — that regularity is the confounder’s signature.)
A is correct — the groups differed before the feature existed: committed customers opt in, so their lower churn may reflect commitment, not the feature. (If you picked B: the data can be flawless and the conclusion still wrong — that’s the whole trap. If you picked C: churn is measurable for both groups; the problem is what the comparison means.)
B is correct — randomizing who gets the feature makes the groups alike except for the feature, so a churn gap can finally be attributed to it; that’s the A/B test, coming in module 9. (If you picked A: opinions from self-selected fans compound the bias. If you picked C: a correlation that persists is still just a persistent correlation.)