The short answer
What is a confounding variable?
A confounding variable, or confounder, is a third factor that influences both of two things that appear related. It makes them rise and fall together without either causing the other. Temperature is the classic confounder behind ice cream sales and drownings: hot weather drives both, so the two track each other even though ice cream does not cause drowning.
The question that saves you
What else could be moving both of these at once?
When two things rise and fall together, the eye jumps to “one must cause the other”. But a correlation has three other explanations before that one: a confounder — some third thing driving both; reversed cause and effect; or sheer coincidence. A correlation, however strong, can never tell these apart on its own. The defence is to hunt for the lurking variable and then compare like with like.
AskIs a hidden third thing driving both — and does the link survive when I hold it still?
01 · The lurking variable
How a third thing fakes a link
The ice-cream-and-drownings story is the textbook case, and it is true: over a year, the two really do track each other closely. The naive reading is a horror film — ban the ninety-nines and save lives. The real story is a weather report. When summer arrives, people buy more ice cream and they swim more, so more of them drown. Heat is pulling both strings at once. Neither has any hold on the other; they are siblings, not parent and child.
That hidden third factor is a confounder, or lurking variable: something that influences both of the things you are comparing, manufacturing a correlation with no causal thread between them. And crucially, this correlation is not a fluke. It is real, strong and stable; collect ten more years of data and it will still be there. What makes it an illusion is not that the numbers lie, but that the obvious explanation — ice cream causes drowning — is the wrong one.
Confounding is the commonest of the reasons a correlation is not causation, and the engine behind Simpson’s paradox too: there, a lurking variable doesn’t just fake a link, it can flip the sign of one. Same disease, different symptom.
02 · Comparing like with like
The move that unmasks the confounder
There is a simple test, and you ran it with the toggle above: hold the suspected confounder still and see if the link survives. Don’t compare a freezing January day with a sweltering August one — of course those differ in everything. Compare hot days only with other hot days. Within a single band of temperature, the dots scatter every which way: on two equally hot days, the one that sold more ice cream is no more likely to have seen more drownings. The slope is gone. The whole upward trend was the temperature talking.
This — comparing like with like, or “controlling for” the third variable — is the backbone of honest statistics. You can do it by splitting the data into matched groups (as the colours do), by adding the confounder to a model so its pull is accounted for, or, best of all, by running an experiment that severs the confounder’s grip entirely. Which is the next move.
03 · Breaking the link
Why randomised trials are the gold standard
The trouble with confounders is that you can only ever control for the ones you think of. There is always the dread of a lurking variable you never measured — and you cannot adjust for a ghost you can’t see.
A randomised controlled trial cuts that knot. Assign who gets the treatment by the flip of a coin, and chance evenly scatters every confounder — the ones you know, the ones you don’t, and the ones nobody has named — across both groups. Now if an effect appears, the treatment is the only thing that systematically differs, so it earns the right to be called the cause. It is exactly why “people who take vitamin X live longer” (the vitamin-takers are richer, more health-conscious, see doctors more — confounders all) so often collapses when someone finally runs the trial.
When a trial is impossible — you can’t randomise people to smoke — careful observational work can still build a case, but only by hunting down confounders one by one and arguing that none big enough remains. That is slow, humble, contestable work, and it is the honest kind.
04 · How not to be fooled
Reading a correlation safely
Name a third cause before you accept the first. For any “A is linked to B” headline, take ten seconds to ask what could drive both — age, wealth, season, health, the kind of person who does A. The candidate is usually not hard to find.
Ask whether they compared like with like. A credible claim has already held the obvious confounders still — “adjusting for age and income”. If a study didn’t, treat its correlation as a question, not an answer.
Check the direction. Maybe B causes A. Depressed people exercise less — or does less exercise cause depression? A bare correlation is silent on which way the arrow points.
Prize the experiment. A randomised trial outranks any number of observational correlations, because randomising is the one move that disarms the confounders you forgot — and the ones you never knew to look for.
Separate this from coincidence. A confounded link is real and won’t wash out with more data; a coincidental one is chance and will. Both are correlation without causation — for opposite reasons.