Zellner (1962) An Efficient Method of Estimating Seemingly Unrelated Regressions and Tests for Aggregation Bias

Summary

Zellner (1962) introduces the Seemingly Unrelated Regression (SUR) model — $M$ regression equations with potentially different regressors but correlated disturbances. A feasible two-stage Aitken Generalized Least Squares (GLS) estimator is proposed: Ordinary Least Squares (OLS) residuals from Stage 1 estimate the cross-equation error covariance $\hat\Sigma_\varepsilon$, which feeds into a single GLS step yielding $b = (X'\hat\Sigma^{-1}X)^{-1}X'\hat\Sigma^{-1}y$; this has the same asymptotic distribution as infeasible GLS. The estimator collapses to OLS when errors are uncorrelated or all equations share identical regressors. The paper also proposes $F$-tests for aggregation bias by testing parameter equality across micro equations, and applies both results to Grunfeld's General Electric (GE)/Westinghouse investment data.

Key Claims

• SUR model: $y_\mu = X_\mu\beta_\mu + u_\mu$, $\mu=1,\ldots,M$, with $E[u_\mu u_{\mu'}'] = \sigma_{\mu\mu'}I_T$ — "seemingly unrelated" because each equation has its own regressors yet correlated errors create informational linkages.
• Two-stage Aitken estimator (eq. 2.10): Stage 1 OLS residuals $\hat{u}_\mu$ $\to$ $(T-l)\hat{s}_{\mu\mu'} = \hat{u}_\mu'\hat{u}_{\mu'}$ $\to$ Stage 2 $b = (X'\hat{S}^{-1}X)^{-1}X'\hat{S}^{-1}y$; asymptotically equivalent to infeasible GLS $b^*$; variance $V(b) = (X'\Sigma^{-1}X)^{-1} + o(T^{-1})$.
• Efficiency conditions: SUR $\equiv$ OLS iff (i) $\sigma_{\mu\mu'}=0$ for all $\mu\neq\mu'$ (uncorrelated errors) or (ii) $X_1=X_2=\cdots=X_M$ (identical regressors). Maximum gain when regressors are orthogonal across equations and errors highly correlated.
• Efficiency gain formula (eq. 3.8): For two equations, $|V(b_1^*)| = \frac{(1-\rho^2)^{l_1}}{\prod_{\mu=1}^{l_1}(1-\rho^2 r_\mu^2)}|\sigma^2(X_1'X_1)^{-1}|$, where $r_\mu$ are canonical correlations between $X_1$ and $X_2$; ratio $\to (1-\rho^2)$ as $r_\mu\to 0$.
• Micro aggregation bias test (eq. 4.4): $H_0:\beta_1=\cdots=\beta_M$; $F_{q,n-m}$ statistic with $q=(M-1)l$ restrictions; asymptotically $\chi^2_q/q$.
• Macro aggregation test (eq. 4.8–4.11): Regress aggregate $\bar{y}$ on $\bar{x}$ and a residual term $w(t)\bar{x}(t)$; non-zero coefficient on the latter indicates aggregation bias.
• Grunfeld empirical application: GE and Westinghouse investment, $T=20$ (1935–1954), regressors $C_{t-1}$, $F_{t-1}$, constant; $\hat\rho(u_1,u_2)=0.81$; SUR variance $\approx 20\%$ below OLS; $\tilde{F}=3.452 > F_{3,34}(0.95)=2.88$, rejecting $\beta_1=\beta_2$ at 5%.

Concepts Introduced or Extended

• Seemingly Unrelated Regression

Entities Mentioned

• Arnold Zellner

Quotes

"...the equations are seemingly unrelated in the sense that each equation has its own set of independent variables."

"...the two-stage Aitken estimate...has the same asymptotic distribution as the Aitken estimate."

My Take

This is the foundational SUR paper. The two-stage Aitken estimator is elegant and practically important — it makes feasible GLS operational without any distributional assumptions beyond the second-moment structure. The two efficiency conditions (uncorrelated errors or identical regressors) are intuitive and form the backbone of all subsequent SUR intuition. The aggregation bias tests are rigorous but assume linear aggregation; the Grunfeld application with $T=20$ makes it easy to illustrate efficiency gains but understates small-sample uncertainty. Zellner (1963) subsequently provided exact finite-sample results under normality.