Description
Estimate several regressions jointly when their errors are correlated across equations at the same unit or occasion, borrowing strength to sharpen each equation's coefficients.
Several outcomes are often measured on the same units, each with its own predictors: time on site and spend per user, or engagement in two markets over the same weeks. Fitted separately, each equation ignores information in the others' residuals. SUR stacks the equations and estimates them by generalised least squares with a cross-equation error covariance estimated from OLS residuals. Zellner's restricted estimator uses each equation's own residuals; the unrestricted one uses residuals from regressing every outcome on all regressors in the system. Gains grow with the cross-equation residual correlation and disappear when every equation has the same regressors, where SUR equals equation-by-equation OLS. Kariya and Kurata's CO2 example shows that misspecified equations inflate the residual correlation, and that revising the mean models can make OLS nearly as efficient.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Write each equation with its own regressors and confirm they share units or occasions so that errors can correlate.
- Fit each equation by OLS and keep the residuals.
- Estimate the cross-equation error covariance from the residuals (restricted or unrestricted Zellner form).
- Check the residual correlations: if near zero, or if all equations share the same regressors, OLS is already efficient.
- Compute the GLS estimate on the stacked system and check each equation for serial correlation and misspecification.
- Report coefficients per equation with SUR standard errors alongside OLS.
Sources
- Generalized Least Squares (Kariya & Kurata, 2004)
Also known as
- SUR
- SURE
- Zellner's estimator
- Restricted Zellner estimator
- Unrestricted Zellner estimator
- Multi-equation GLS
Categories
- Analytics & Behaviour
- Statistics & Modelling