Statistics Glossary → Seemingly unrelated regression (SUR) and 3SLS
Seemingly unrelated regression (SUR) and 3SLSconcept
System methods that estimate several regression equations, each with its own outcome, jointly by exploiting the correlation between the equations' errors.
In Zellner's SUR model each equation has its own outcome and predictors; the equations are 'seemingly' unrelated, but because they are measured on the same units their errors can be correlated (e.g. shared unobserved factors affecting a family's food and clothing spending). SUR first estimates each equation by ordinary least squares, computes the cross-equation error covariance matrix from the residuals, and then re-estimates the whole system by feasible generalized least squares (FGLS). Three-stage least squares (3SLS) applies the same idea to simultaneous equation systems with endogenous regressors: the first two stages are 2SLS equation by equation, and the third is system-level GLS.
When errors are correlated across equations, system estimation is more efficient than estimating the equations separately (standard errors shrink). If the errors are uncorrelated, or all equations contain exactly the same predictors, SUR gives the same result as equation-by-equation least squares. 3SLS has a price: a specification error in one equation spreads to the others through the system estimation, which makes single-equation methods (2SLS) more robust in this respect. The choice of denominator for the error covariance (T or degrees-of-freedom corrected) and of iterated estimation can change the results slightly.
An everyday example
A researcher links families' food and transport spending to income and household size in two separate equations. Families hit by an economic shock cut back on both items together, so the two equations' errors are correlated (r = 0.45). SUR uses this correlation to give more precise coefficients than estimating the two equations separately.
In MerQur
Available as the system options of 'Estimation type' in instrumental variables regression: SUR (seemingly unrelated regression) and 3SLS (three-stage least squares). The 'System equations', 'Residual covariance denominator' and 'Iterated (ITSUR / IT3SLS)' options define the system and the details of estimation.
Analyses where this concept appears
Instrumental Variables Regression (2SLS)