A regression of salary on education may combine the effect of schooling with characteristics related to both schooling and earnings. Leaving out a relevant variable does not always bias the slope: the omitted variable must affect the outcome and correlate with the included regressor.
Two models, two slopes
Suppose the data-generating mechanism is:
represents education and an invented latent ability variable. If we fit only , the population projection slope is:
For positive and positive correlation, the simple regression attributes part of the effect of to education, inflating the slope. Setting correlation to zero removes this term; negative correlation reverses its sign.
- Simple slope α̂₁
- —
- Controlled slope β̂₁
- —
- β̂₂
- —
- Cov(X₁, X₂)
- —
- Var(X₁)
- —
- Bias = α̂₁ − β̂₁
- —
Simulation: Y = 10 + 2X₁ + 8X₂ + ε. Ability is an invented standardized variable, not an observed biological trait. All three controlled lines are conditional fits from one multiple regression.
What controlling means
The controlled regression estimates both coefficients simultaneously. The three lines are conditional predictions at three values of : they share an education slope but have different intercepts. They are not three independent regressions.
The simulation fixes and for 90 observations. Marginal dispersions remain constant as correlation changes. Disturbances are constructed orthogonal to the regressors so the decomposition also holds exactly in this sample; real datasets add sampling variation. The panel computes the simple and multiple fits rather than merely drawing preset slopes.
Methodological conclusion. Controlling for a covariate does not automatically establish causality. Measurement, causal structure and remaining omitted variables all matter. Ability here is a teaching device, not an empirical measure of innate differences or a definitive account of wage inequality.