Statistics Glossary → Inverse Mills ratio
Inverse Mills ratioconcept
The ratio of the standard normal density to the cumulative distribution, λ(z) = φ(z) / Φ(z); the term added to a regression in the Heckman selection model to correct for bias caused by sample selection.
When the outcome is observed only in a selected subgroup (e.g. wages only for those in work), ordinary regression is biased if the selection process is related to the outcome. In Heckman's two-step method, the probability of selection is first estimated with a probit model, and the inverse Mills ratio is computed from z'γ for each selected observation. This ratio is then added to the outcome equation as an extra regressor, thereby modelling the non-zero conditional mean of the error term in the selected sample.
The coefficient of λ equals the correlation between the errors of the selection and outcome equations (ρ) times the standard deviation of the outcome error (σ); a significant coefficient points to selection bias. It is advisable for the selection equation to contain at least one variable absent from the outcome equation (an exclusion restriction); otherwise λ is nearly collinear with the other regressors, identification rests only on the nonlinearity of the probit function, and the results become unstable. The method is sensitive to the assumption that the errors are bivariate normal.
An everyday example
In a model explaining women's wages by education, wages are observed only for women in work. Because some women with young children who receive low wage offers choose not to work, the wages of those working do not represent the potential wages of all women. Adding the inverse Mills ratio derived from the decision to work corrects for this selection effect.
In MerQur
In Sample Selection and Truncated Regression, the 'Save the inverse Mills ratio to the dataset' checkbox adds λ = φ(z'γ) / Φ(z'γ) for the selected rows as a new column.
Analyses where this concept appears
Sample Selection and Truncated Regression (Heckman)