Omitted Variable Bias

Also known as · OVB

Bias in β^1\hat\beta_1 when a relevant variable x2x_2 correlated with the included x1x_1 is left out: it folds into the error, violating A2, and the bias is β~2S12/S11\tilde\beta_2 S_{12}/S_{11}. There is no bias if S12=0S_{12}=0 or if the omitted variable's true coefficient is zero.

When to use

Use the sign table (true effect times covariance) to say whether the estimate is biased up or down. Worked cases are the computer-price Hedonic Pricing example and the birthweight regression in Assignment 4, where the cigarette coefficient moved from −10.98-10.98 with controls to −12.906-12.906 without.

See Lecture 9 and Lecture 13 and Lecture 11.

Appears in