MerQuR+Statistics Glossary

Statistics Glossary → Sample selection bias (Heckman)

Sample selection bias (Heckman)concept

Bias in estimates that arises because the outcome is observed only for units that pass through a selection process related to the outcome; the Heckman model was developed to correct it.

The classic example is the wage equation: wages are observed only for people who work, and the decision to work is related to unobserved traits (motivation, expected wage). A regression on workers alone therefore gives biased coefficients for the whole population. The Heckman (1979) model has two equations: a probit selection equation for the probability that a unit is observed, and an outcome equation for y, with ρ the correlation between their errors. If ρ = 0 there is no selection bias. In the two-step (Heckit) method, the inverse Mills ratio computed from the probit estimates is added to the outcome equation, while maximum likelihood estimates both equations together.

Reliable estimation needs at least one variable that appears in the selection equation but does not directly affect the outcome (an exclusion restriction), such as the number of young children in a wage model. Without one, the model is identified only through the nonlinearity implied by the normality assumption, and the results can be very unstable. The model is sensitive to the assumption that the errors are bivariate normal. Truncated regression is a different situation: there, units outside the limit never enter the sample, and selection depends on a threshold in y itself.

An everyday example

A researcher wants to explain women's wages by education, but wages are observed only for women in employment. If women who expect a low wage are less likely to work, those in employment are selected as 'able to earn high wages', and the effect of education is estimated wrongly. The Heckman model corrects for this selection, using the number of children as the exclusion variable.

In MerQur

In Sample Selection and Truncated Regression, 'Selection equation variables' specifies the probit selection equation (an exclusion restriction is recommended), the 'Model' option chooses Heckman maximum likelihood, two-step (Heckit) or truncated regression, and the 'Selection indicator' specifies which units were observed.

Analyses where this concept appears

Sample Selection and Truncated Regression (Heckman)

Related concepts

Inverse Mills ratioTruncation