What is a Confounding Variable?

A confounding variable is a hidden factor that pushes on two things at once, so both change together and look connected to each other when they are only connected to the confounding variable. The pattern is real; the explanation is wrong. Coffee drinkers once appeared to have more heart disease, until researchers saw that they also smoked more, and smoking was doing the damage. The confounder hid outside the two things measured, which is where confounders always hide.

A confounding variable is a factor outside a study that influences both the supposed cause and the outcome, producing a correlation between them that owes nothing to one causing the other. Coffee drinking was long linked to heart disease. Coffee drinkers also smoked more than non-drinkers, and smoking causes heart disease. Smoking was the confounder: tied to coffee through habit, tied to heart disease through biology, and enough to make coffee appear responsible. The test for a confounder is that it must connect to both ends. A factor linked only to the cause or only to the outcome cannot manufacture a link between them. Confounders are dangerous because they sit outside the data unless someone thought to measure them. A study that recorded only coffee and heart health would never see smoking. Researchers handle them three ways. They can exclude people who carry the confounder, studying only non-smokers. They can randomise, assigning treatment by chance so smokers spread evenly across groups. Or they can measure the confounder and adjust for it, comparing coffee drinkers with non-drinkers who smoke the same amount. The first two must be designed in from the start. The third works only for confounders somebody named.

Measure a school full of children and shoe size predicts reading ability almost perfectly: the bigger the feet, the better the reader. Nobody concludes that big feet help you read. Age is doing both jobs. Older children have bigger feet and older children read better, so the two rise together while neither does anything to the other. Age is the confounder: it sits behind both measurements, it was never one of the two things compared, and it made the comparison meaningless. Compare only the eight-year-olds with each other and the link between feet and reading vanishes, which is what removing a confounder looks like.