Arnold Zellner

Overview

Arnold Zellner (1927–2010) was a Bayesian econometrician at the University of Chicago Booth School of Business. He introduced the Seemingly Unrelated Regression (SUR) model (1962) and the GLS estimator that exploits cross-equation error correlation. His textbook An Introduction to Bayesian Inference in Econometrics (1971) was foundational. He also developed the g-prior for regression coefficients, contributed to the triangular reparameterization of simultaneous equation systems (Zellner-Bauwens-Van Dijk 1988), and in his 1985 Fischer-Schultz Lecture argued that prior information is unavoidable in econometrics and demonstrated the severe overparameterization of unrestricted VARs.

Key Contributions / Features

• Tiao-Zellner (1964): When two regression samples share slope vector β but have independent unknown variances σ₁² and σ₂², the Bayesian posterior of β under noninformative priors is a product of two multivariate t-distributions — the "double-t" distribution. Normalizing constant is a p-dimensional integral; asymptotic expansion in ν₁⁻¹ and ν₂⁻¹ makes it tractable. Theil's (1963) mixed estimator emerges as the Bayesian limit when ν₁ → ∞ with σ₁² known. Biometrika 51(1/2): 219–230. See Tiao-Zellner (1964).
• Zellner (1962): Introduced the SUR model — $M$ regression equations $y_\mu=X_\mu\beta_\mu+u_\mu$ with cross-correlated errors. Derived the two-stage Aitken (feasible GLS) estimator asymptotically equivalent to infeasible GLS; showed efficiency over OLS is zero iff errors are uncorrelated or all $X_\mu$ are equal; expressed the gain via canonical correlations between regressors. Proposed $F$-tests for aggregation bias. Grunfeld investment application: $\hat\rho=0.81$, ~20% variance reduction. See Zellner (1962).
• Zellner (1976): "Bayesian and Non-Bayesian Analysis of the Regression Model with Multivariate Student-t Error Terms" (JASA 71(354): 400–405) — showed that OLS is MLE and MVLUE under multivariate Student-t errors; classical t/F statistics remain marginally valid; scale inference requires F rather than χ²; marginal posterior for β under diffuse prior is invariant to ν₀ (same form as normal model). Scale-mixture-of-normals representation anticipates MCMC-based robust Bayesian regression. See Zellner (1976).
• Zellner (1971): An Introduction to Bayesian Inference in Econometrics — standard reference for Bayesian multivariate regression, SUR posteriors under diffuse and conjugate priors.
• Zellner (1985): "Bayesian Econometrics" (Econometrica 53(2): 253–270) — Fischer-Schultz Lecture establishing Bayesian methods as the natural language of econometrics. Key arguments: (1) a 6-variable MVARMA(3,4) has 395 parameters vs. ≈480 observations (obs/param ≈ 1.2), making reliable unrestricted estimation infeasible; (2) VAR(10) with 6 variables has 387 parameters; (3) the 3-equation Demand-Supply-Entry structural model has only ≈20 parameters, demonstrating how theory creates parsimony; (4) Bayesian pre-test estimate $\hat\theta^* = \theta_0 + (1+K_{12})^{-1}(\bar\theta_2-\theta_0)$ interpolates between null and unrestricted estimator; (5) optimal policy $x^* = (y_*/E\beta)/[1+\operatorname{var}\beta/(E\beta)^2]$ — parameter uncertainty induces caution. See Zellner (1985).
• Zellner (1988): "Optimal Information Processing and Bayes's Theorem" (The American Statistician 42(4): 278–280) — derives Bayes's theorem as the unique optimal information processing rule (IPR) via calculus of variations; defines the Information Conservation Principle (ICP: output information = input information); shows the Bayesian IPR is 100% efficient ($\Delta[\pi_p^*]=0$); connects Bayesian updating to Kullback-Leibler divergence minimization and maximum-entropy principles. Published with discussions by Jaynes, Hill, Bernardo, and Kullback. See Zellner (1988).
• Zellner (2002): "Bayesian Shrinkage Estimates and Forecasts of Individual and Total or Aggregate Outcomes" (UChicago Working Paper 0204) — analyzes whether shrinkage of individual parameters/forecasts improves estimation of their aggregate; key distinction: Stein's zero-mean prior shrinks totals, Lindley's free-mean extension does not (deviations shrunk, level preserved); BMOM conceptual-sample approach yields shrinkage total estimate without likelihood/prior; Lemmas 1–2 give sufficient conditions (diagonal MSE matrices) for shrinkage total estimator to dominate; motivated by Zellner-Chen (2001) 11-sector GDP forecasting application. See Zellner (2002).
• Zellner-Bauwens-Van Dijk (1988): Triangular reparameterization of the SUR system that orthogonalizes the parameter blocks, enabling closed-form Monte Carlo inference; later exploited by Zellner-Ando (2010).
• Zellner and Ando (2010): Direct Monte Carlo (DMC) approach for Bayesian SUR inference; shows Gibbs sampling fails in high-dimensional SUR (92/100 convergence failures when $p_j=25$, $n=50$); DMC generates independent draws with no burn-in. See Zellner and Ando (2010).
• Bayesian Method of Moments (BMOM) (Zellner 1994, 1997a; Zellner-Tobias-Ryu 1997): constructs post-data densities for parameters without specifying a likelihood — uses maximum entropy subject to moment constraints (e.g., E[β|D] = OLS, Cov[β|D] = OLS covariance). For standard regression, MaxEnt posterior coincides with the diffuse-prior Bayesian Normal-t result but needs no distributional assumption. Extended to semiparametric regression via series expansions (Polynomial, Fourier, Gallant), with model-selection via estimation loss, predictive loss, and posterior odds. BMOM with conceptual sample yields Stein-type shrinkage estimators (Zellner 2002). See Bayesian Method of Moments and Zellner-Tobias-Ryu (1997).
• SEMTSA and ARLIWI (Zellner-Palm 1974/75; Zellner c.2000): Structural Econometric-Time Series Analysis links structural economic theory (demand-supply-entry, IS-LM, RBC) to implied ARMA/VAR forecasting equations. The ARLIWI model Δy_t = f(lags(Δy), ΔSR_{t-1}, Δm_{t-1}, world income) applied to 18 industrialized countries achieves pooled RMSE 1.74% and ~70% turning-point accuracy in 211 episodes. KISS principle: "Keep It Sophisticatedly Simple" — simple theory-consistent models outperform complicated macroeconometric models. See SEMTSA and Zellner (c.2000).
• Marshallian Macroeconomic Model (MMM) (Zellner-Chen 2000): structural disaggregated macro forecasting model; demand + supply + entry equations per sector solve to logistic differential equation (1/S)dS/dt = a(1−S/F)+g; 11 U.S. sectors summed for GDP forecast; disaggregated MMM(DA)III + SUR achieves RMSE=1.40 vs AR(3) RMSE=2.26 for 1980–1997; currency beats M1 as leading indicator; Extended MELO balanced-loss estimator (L=w·fit²+(1−w)·precision²); direct Monte Carlo for finite-sample posterior/predictive densities. See Zellner-Chen (2000).
• Zellner-Min (1995): Three operational convergence criteria (ARC², DC², RC²) for the Gibbs sampler that detect convergence to incorrect distributions — exploiting the identity p(α|D)p(β|α,D) = p(β|D)p(α|β,D) as a computable check on GS output. See Zellner-Min (1995).

Related

• Bayesian Linear Regression
• Seemingly Unrelated Regression
• George C. Tiao
• Tiao-Zellner (1964)
• Gibbs Sampler
• Tomohiro Ando
• Vector Autoregression
• Minnesota Prior
• Heavy-Tailed Distributions
• Zellner (1976)
• Zellner (1985)
• Bayesian Method of Moments
• Structural Econometric-Time Series Analysis (SEMTSA)