Correlation, Causation & Confounding
P1.stats-literacy.01 · Audience: guest, it-ml, language-pro
Take it one idea at a time: why two numbers moving together doesn't mean one causes the other, why an aggregate can lie, why a rebound isn't proof your fix worked, and why busy dashboards throw up coincidences. Each step has a worked example. No maths required.
One mental model. Before you act on a pattern in the data, ask "what else could explain this?" Most bad decisions in this module come from skipping that one question.
A correlation just means two things tend to move together. It does not mean one causes the other. A hidden third factor — a confounder — can quietly drive both.
Worked example — the greeter that 'boosted' sales
A retailer notices stores with a door greeter sell more, and nearly rolls greeters out everywhere. Then someone asks what else could explain this?
| Store type | Has a greeter? | Monthly sales |
|---|---|---|
| Flagship (big, busy) | yes | high |
| Local (small, quiet) | no | lower |
Greeters were only ever assigned to the big flagship stores — which already sold more for reasons of size and footfall. Store size is the confounder: it drives both the greeter and the sales. The greeter may do nothing at all.
The gold standard for proving cause is a controlled experiment (an A/B test): add a greeter to a random half of comparable stores, hold the rest steady, and compare.
⚡ Interview Ref — the quick-scan Reference face
In plain language: why two numbers moving together doesn't mean one causes the other, why an aggregate can lie, why a rebound isn't proof your fix worked, and why busy dashboards throw up coincidences. No maths required.
One mental model. Before you act on a pattern in the data, ask "what else could explain this?" Most bad decisions on this page come from skipping that one question.
1 — Correlation vs causation for business decisions (b1)
A correlation just means two things tend to move together. It does not mean one causes the other. A hidden third factor — a confounder — can quietly drive both.
The classic example: ice-cream sales and drownings both rise together. Ice cream doesn't cause drowning — summer heat drives both. Heat is the confounder.
In a business setting the same trap appears everywhere:
- Teams using the new tool ship faster — or are the fast teams just the early adopters?
- Newsletter readers buy more — or do keen buyers simply sign up for more emails?
Before you act, ask what else could explain the link. The gold standard for proving cause is a controlled experiment (an A/B test): change one thing, hold the rest steady, and compare.
2 — Simpson's paradox (b7)
Simpson's paradox is when a trend that shows up in every subgroup reverses once you pool the groups together — because the mix of groups is unbalanced.
A plain example: a treatment looks worse overall than the alternative, yet it's actually better for mild cases and better for severe cases. How? It was given mostly to the sicker patients, who do worse no matter what — so the tough cases drag it down.
The same shape shows up in business:
- A channel looks weak overall but wins in every region — it's over-weighted to the hardest region.
- Team B's average score trails Team A's, yet B beats A in both junior and senior roles.
Lesson: always check whether an aggregate hides an imbalanced segment mix. Break the number down by segment before you trust the headline.
3 — Regression to the mean (b8)
Regression to the mean is simple: extreme results tend to be followed by more average ones, purely by chance. A record-bad month is usually part luck, and luck doesn't repeat, so the next month drifts back toward normal on its own.
This quietly hands false credit to interventions. You launch a fix right after a terrible month, the numbers bounce back — and the fix takes the applause for a rebound that would have happened anyway.
Everyday cases:
- Shout at the worst reps and they improve next quarter — but so would anyone at a low ebb.
- A store's worst week is followed by a better one whether or not you changed the display.
The fix: use a control group or a proper baseline. Compare the treated group against an untouched one over the same period, so a natural rebound can't masquerade as impact.
4 — Spurious correlation; not being fooled by dashboards (b11)
With enough metrics on a dashboard, some pairs will line up by pure chance. That's a spurious correlation — a coincidence with no real link behind it.
The maths is unforgiving: the more pairs you test, the more accidental matches you'll find. Test a hundred metrics against each other and a few will look 'significant' for no reason at all — that's how testing many things inflates false findings.
How not to get fooled:
- Demand a plausible mechanism — a story for why one thing would move the other.
- Be wary of cherry-picked pairs — anyone hunting a dashboard will find one that fits.
- Pre-register the question — decide what you expect to see before you go looking.
A striking chart is a starting question, not a conclusion.
Interview one-liners
- Correlation isn't causation — look for the hidden confounder, and prove cause with a controlled experiment.
- Simpson's paradox — an aggregate can reverse a trend that holds in every segment; always split by segment.
- Regression to the mean — extremes drift back on their own, so use a control group before crediting a fix.
- Spurious correlations — busy dashboards throw up coincidences; demand a mechanism, not a matching chart.
📚 Go Further
Plain-language reads on reasoning about cause and evidence.
| Type | Resource |
|---|---|
| Book | Thinking, Fast and Slow — Kahneman, on regression to the mean and bias |
| Book | The Book of Why — Judea Pearl, on correlation, confounding, and cause |
| Site | Spurious Correlations — Tyler Vigen, a gallery of coincidental pairs |
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
Where next?
Later in Data & Statistics Literacy