Zellner (1985) Bayesian Econometrics

Summary

Zellner's 1985 Fischer-Schultz Lecture — published in Econometrica 53(2): 253–270 — makes the programmatic case that Bayesian methods are the natural language for econometrics: prior information is unavoidable (implicit in model choice, identification, lag truncation, and variable selection), so it should be made explicit and formal. He demonstrates through parameter-counting that unrestricted Multivariate AutoRegressive Moving Average (MVARMA) and Vector Autoregression (VAR) models are severely overparameterized relative to available data (395 parameters for a 6-variable MVARMA(3,4) vs. $\approx$480 observations; 387 parameters for an unrestricted VAR(10) with 6 variables), and proposes the 3-equation Demand-Supply-Entry (DSE) structural model as a parsimonious alternative with only $\approx$20 free parameters. The paper surveys Bayesian inference — posterior, marginal, and predictive densities; posterior odds and Bayesian pre-test estimators; and optimal policy control under parameter uncertainty — illustrating each with applied evidence from Litterman's Bayesian VAR (BVAR) through Bayesian portfolio choice.

Key Claims

• Prior information is always present in applied econometrics, whether recognized formally or not; the choice of model dimension, identification restrictions, lag length, and variable set all embed prior beliefs.
• A 6-variable MVARMA($r=3$, $q=4$) system has 395 parameters against $\approx$480 observations (obs/param $\approx$ 1.2); reliable estimation is infeasible without informative priors.
• An unrestricted VAR(10) with 6 variables has 387 parameters; VAR(4) has 159 — both are heavily overparameterized relative to typical macroeconomic samples.
• The DSE 3-equation competitive industry model (demand, supply, entry) has $\approx$20 structural parameters vs. the 144-element VAR parameter matrix; structural theory creates parsimony that data alone cannot.
• Bayesian posterior: $p(\theta|D) = c\,p(\theta|I_0)\,p(y|\theta)$ — a finite-sample exact result requiring no asymptotics.
• Marginal posterior integrates out nuisance parameters: $p(\theta_1|D) = \int p(\theta_1,\theta_2|D)\,d\theta_2$.
• Predictive probability density function (pdf): $p(y_f|D) = \int p(y_f|\theta)\,p(\theta|D)\,d\theta$ — averages over parameter uncertainty rather than plugging in point estimates.
• Posterior odds: $K_{12} = (\Pi_1/\Pi_2)\times[p_1(y|H_1)/p_2(y|H_2)]$ — combines prior odds with Bayes factor; avoids pre-test distortions.
• Bayesian pre-test estimate (Stein-like shrinkage): $\hat\theta^* = \theta_0 + (1+K_{12})^{-1}(\bar\theta_2 - \theta_0)$; interpolates between null and unrestricted estimator with weight determined by posterior odds.
• Optimal policy control: $x^* = (y_*/E\beta)/[1+\operatorname{var}\beta/(E\beta)^2]$; parameter uncertainty in the multiplier $\beta$ reduces the optimal policy instrument value — formal Bayesian argument for cautious policy.
• Applied Bayesian results (Litterman BVAR, Monahan AutoRegressive Integrated Moving Average (ARIMA), Griffiths-Dao, Fomby-Guilkey, Swamy-Mehta, Bawa-Brown-Klein/Jorion portfolio) systematically outperform classical counterparts in forecasting accuracy and statistical efficiency.

Concepts Introduced or Extended

• Vector Autoregression (parameter proliferation critique; Bayesian shrinkage as remedy)
• Minnesota Prior (cited as primary applied Bayesian VAR solution)
• Bayesian Hierarchical Model (posterior, marginal, predictive density framework)
• Marginal Data Density (posterior odds / Bayes factor for model comparison)

Entities Mentioned

• Arnold Zellner
• Robert Litterman (BVAR shrinkage prior as leading Bayesian VAR example)

Quotes

"The main point is that prior information is always present in applied econometric work whether recognized formally or not."

"It is clearly desirable to have approaches to econometric estimation and testing that employ prior information in a flexible, formal manner and thus avoid some of the problems of data mining, pre-testing, and related practices."

"Those who advocate just using a large VAR with many variables and many lags have not performed the calculations [to see] that the number of observations available is barely sufficient."

My Take

This is primarily a programmatic manifesto rather than a technical contribution, yet its influence has been significant: the parameter-counting critique of large VARs anticipated the renewed interest in BVAR methods (Litterman, Sims-Zha), and the DSE parsimony argument foreshadows modern Dynamic Stochastic General Equilibrium (DSGE)-VAR debates. The pre-test estimator formula is elegant and under-appreciated; the optimal control result formalizes the intuition that uncertainty should induce cautious policy, connecting naturally to Hansen-Sargent robust control. The paper's main limitation is that "applied examples" are surveyed rather than reproduced — results are reported in summary form from other authors' work, making independent verification difficult. Zellner's framing of prior information as inevitable rather than optional remains the strongest argument for Bayesian econometrics.