Omitted variable bias
Leaving a relevant cause out of a regression warps the coefficients you keep.
What it means
Omitted variable bias is the error in an estimated relationship that results when a variable correlated with both an included predictor and the outcome is left out of the model. The omitted cause's influence gets absorbed into the coefficients of the included variables, biasing them up or down depending on the directions of correlation. It is the regression-specific face of confounding, and it cannot be diagnosed from goodness-of-fit alone. Economists rely on its sign-and-magnitude logic to reason about whether a naive estimate overstates or understates the true effect.
Examples
Estimating the wage return to schooling without controlling for ability overstates schooling's effect, since abler people both study more and earn more.
A study finds that people who take vitamins live longer, but leaves out income. Wealthier people buy supplements and also eat well and see doctors, so the pills collect credit they never earned.
A retailer's model shows email campaigns lifting sales, but omits the season. The emails go out in December, so the coefficient on email quietly swallows Christmas and the team doubles the send rate.
First described in Econometrics; standard since the mid-20th century.