Zellner (2000) Bayesian and Non-Bayesian Approaches to Scientific Modeling and Inference in Economics and Econometrics

Summary

Invited address at Ajou University (Korea) in honor of Professor Tong Hun Lee, written circa 1999–2000. Surveys Zellner's integrated research program across Bayesian methodology, information theory, and structural macroeconomic forecasting. Argues the Bayesian approach is superior to frequentist methods on both philosophical and practical grounds; connects the Bayesian method of moments (BMOM) / maximum-entropy (maxent) approach to the Bayesian framework; and presents the Structural Econometric-Time Series Analysis (SEMTSA)–ARLIWI forecasting program applied to 18 industrialized countries. The KISS principle ("Keep It Sophisticatedly Simple") ties the methodological and empirical threads together.

Key Claims

• Pearson-Jeffreys "unity of science": "The unity of all science consists alone in its method, not in its material" (Pearson 1938). Any field is scientific if scientific methods are applied; Jeffreys' Theory of Probability provides the operational Bayesian framework applicable across all sciences.
• Table 1 — Bayes vs Non-Bayes: Bayesian approach has (1) a formal learning model (Bayes' theorem), (2) axiomatic support, (3) probabilities assigned to hypotheses and models, (4) a degree-of-confidence interpretation of probability, (5) finite-sample posterior intervals $\Pr(a < \theta < b \mid \text{data})$, (6) finite-sample predictive intervals. Frequentist approach has none of these; p-values are not posterior probabilities despite widespread misinterpretation.
• Bayes' theorem is a 100% efficient information processing rule (Zellner 1988): the unique IPR that conserves information — no other rule achieves the same output information from the same input. Establishes a link between Bayesian updating and information theory (Kullback-Leibler (KL) divergence minimization).
• BMOM / maxent extension: When the likelihood is unknown, maxent densities satisfying moment conditions are the least informative distributions compatible with the data, providing a Bayesian-flavoured analysis without likelihood specification. Dynamic extensions of BMOM differ from the static Bayes' theorem solution, just as dynamic vs static optimization solutions differ in economics.
• ARLIWI forecasting model (autoregressive + leading indicator + world income): $\Delta y_t = \delta_0 + \sum_i \delta_i \Delta y_{t-i} + \delta_4 \Delta SR_{t-1} + \delta_5 \Delta m_{t-1} + \delta_6 \Delta SR_{t-2} + \delta_7 \Delta m_{t-1} + \delta_8 \Delta W_t + \delta_9 \Delta ex_t$. Derived from aggregate supply-demand / Hicksian IS-LM / generalized real business cycle (RBC) models (SEMTSA approach). Applied to 18 industrialized countries 1954–1995; pooled root mean squared error (RMSE) range 1.17–2.53%, median 1.74%; unpooled median higher (~2.5%). Disaggregated forecasting (forecast each country, take median) cuts RMSE by ~20% vs aggregate approach.
• Turning-point forecasting: ~70% correct in 211 turning-point episodes across 18 countries. Methodology: compute predictive density for next year's growth; P = probability of growth below current year = downturn (DT) probability; optimal forecast is DT if $P > \tfrac{1}{2}$ under symmetric loss.
• KISS principle: "Keep It Sophisticatedly Simple" (Zellner's reinterpretation of KISS). Simple models dominate complicated ones in science — Newton's laws, Einstein's laws, demand/supply — while complicated macroeconometric models routinely fail to beat AR benchmarks. Consistent with Jeffreys-Wrinch simplicity postulate.
• Non-Bayesian approaches in practice: mixed estimation (Theil-Goldberger), ridge regression, and random effects models already implicitly introduce prior-like shrinkage, blurring the Bayes/non-Bayes distinction. Good (1991) identifies this as a possible compromise.

Concepts Introduced or Extended

• Bayesian Method of Moments
• Structural Econometric-Time Series Analysis (SEMTSA)
• Bayesian VAR

Entities Mentioned

• Arnold Zellner

Quotes

"IT PAYS TO GO BAYES."

"Keep It Sophisticatedly Simple. Indeed it is hard to find a single complicated model in science that has performed well in explanation and prediction."

"The unity of all science consists alone in its method, not in its material." — Karl Pearson

My Take

A valuable synthesis paper that ties together Zellner's entire research agenda into one coherent vision: Bayesian methods as optimal information processing, BMOM as a likelihood-free extension, SEMTSA as the bridge between theory and time series, and KISS as the unifying methodological principle. The ARLIWI forecasting results remain the most concrete empirical contribution — 1.74% pooled RMSE with theory-grounded regressors, competitive with professional forecasters. The Bayes/non-Bayes comparison table is pedagogically effective and frequently cited in Zellner's work. Limitation: as a conference address, the paper covers breadth at the cost of depth — no new theoretical results, mostly a synthesis and advertisement for existing work. The BMOM/maxent extension to dynamic settings is mentioned but not fully developed.