Loading field map
Bayesian econometrics begins from a simple but radical premise: uncertainty about economic parameters should be represented as a probability distribution, and that distribution should be updated when new data arrive using Bayes' theorem. This commitment sets it apart from the dominant frequentist tradition, which treats parameters as fixed unknowns and evaluates procedures over hypothetical repeated samples. The practical challenge that has driven the subfield's evolution is not philosophical but computational: how to actually compute the posterior distribution when economic models are complex, parameters are numerous, and the integrals involved have no closed form. The history of Bayesian econometrics is therefore a story of five frameworks that each expanded the range of models that could be handled, while also deepening the debate about what kind of prior information is legitimate and how to cope with uncertainty about the model itself.
The first systematic framework, Prior-Based Bayesian Econometrics, was crystallized by Arnold Zellner's 1971 book An Introduction to Bayesian Inference in Econometrics. Zellner showed that Bayesian methods could be applied to standard econometric models—linear regression, simultaneous equations, seemingly unrelated regressions—by choosing conjugate priors that combined analytically with the likelihood to produce a tractable posterior. The distinctive contribution of this framework was not merely that it offered an alternative to frequentist inference, but that it forced practitioners to be explicit about their prior beliefs. Where a frequentist might appeal to asymptotic approximations, the Bayesian could produce exact finite-sample posterior distributions, a major advantage when sample sizes were small. The price of this exactness was severe: only models with conjugate priors were computationally feasible, and the prior itself was typically chosen for mathematical convenience rather than as a faithful representation of genuine economic knowledge. This tension—between the ideal of incorporating substantive prior information and the practical need for tractable functional forms—would become a recurring theme.
The computational bottleneck that constrained Prior-Based Bayesian Econometrics was broken by Simulation-Based Bayesian Econometrics, which emerged in the late 1970s. The key innovation was to replace analytical integration with Monte Carlo simulation. In a landmark 1978 paper, Kloek and van Dijk demonstrated that posterior moments of simultaneous equation models could be estimated by drawing random samples from the posterior density, even when that density had no closed form. This was a transformation of the practitioner's relationship to the prior: because simulation could handle arbitrary prior-likelihood combinations, the prior no longer had to be conjugate. The framework's later development of Markov chain Monte Carlo (MCMC) methods in the 1990s turned simulation into universal infrastructure for Bayesian inference. Simulation-Based Bayesian Econometrics did not replace Prior-Based Bayesian Econometrics so much as absorb it: conjugate priors remained useful as convenient special cases, but the entire field now operated under the assumption that any posterior that could be simulated could be used for inference. The framework's lasting legacy is that it made Bayesian methods applicable to models—nonlinear, non-Gaussian, high-dimensional—that were previously inaccessible to either analytical Bayesian or frequentist approaches.
While simulation methods were expanding the general toolkit, a specific application domain was generating its own framework: Bayesian Time Series and VAR Econometrics. Vector autoregressions (VARs), introduced by Christopher Sims in the early 1980s, promised a flexible way to model macroeconomic dynamics without the restrictive theoretical assumptions of large structural models. But VARs suffered from severe overparameterization: with many variables and lags, the number of free parameters quickly exceeded the available data, producing imprecise estimates and poor forecasts. The Bayesian response, developed by Robert Litterman and colleagues at the Federal Reserve Bank of Minneapolis, was the Minnesota prior. This prior shrinks the VAR coefficients toward a random walk, reflecting the belief that economic time series are persistent but not perfectly predictable. The Minnesota prior is a shrinkage prior, not a theory-based prior: it encodes a statistical regularity rather than a structural economic relationship. This distinction matters because it shows that Bayesian Time Series and VAR Econometrics coexists with Simulation-Based Bayesian Econometrics as an application of its computational tools, while also carving out a distinctive role for priors as regularizers rather than carriers of substantive knowledge. The framework was rapidly adopted by central banks and policy institutions, where it became the standard tool for forecasting and conditional projection.
Hierarchical Bayesian Econometrics, which took shape in the mid-1990s, addressed a different limitation of the earlier frameworks: their treatment of heterogeneity across individuals, firms, or markets. In standard Bayesian analysis, each parameter has its own prior; in a hierarchical model, parameters are drawn from a common population distribution whose own parameters (hyperparameters) are estimated from the data. This partial pooling structure allows information to be shared across units, improving estimates for each unit, especially when individual-level data are sparse. The framework's distinctive contribution is to treat the prior itself as an object of inference: rather than being a fixed encoding of the researcher's beliefs, the prior becomes an estimable distribution that describes the population from which the units are drawn. This conceptual shift connected Bayesian econometrics to the structural microeconometrics tradition, where random coefficients and heterogeneous agent models were already central. Hierarchical Bayesian Econometrics did not replace the earlier frameworks; it layered on top of them, using simulation methods to sample from the high-dimensional posterior and often employing time-series or prior-based submodels as building blocks. Its current role is strongest in microeconometric applications—demand estimation, labor supply, marketing—where individual-level heterogeneity is the object of study rather than a nuisance.
The fifth framework, Bayesian Model Averaging and Model Uncertainty, emerged in the late 1990s from a growing unease with the standard practice of selecting a single model and then conditioning all inference on that choice. Model selection ignores the uncertainty about which model is correct, leading to overconfident estimates and predictions. Bayesian Model Averaging (BMA) addresses this by assigning prior probabilities to a set of candidate models, computing the posterior probability of each model given the data, and then averaging predictions or parameter estimates across models weighted by these posterior probabilities. The mechanics require specifying a prior over the model space—how many variables to include, which functional forms to consider—and this prior can substantially influence the results. The framework's relationship to the other four is that of a meta-layer: BMA operates on top of model classes defined by Prior-Based, Simulation-Based, Time-Series, or Hierarchical frameworks)Skip. It does not replace them; it adds a layer of uncertainty quantification that those frameworks, focused on inference within a given model, had largely ignored. The philosophical tension within BMA—between the ideal of averaging over all plausible models and the practical necessity of restricting the model space—mirrors the broader Bayesian tension between principled uncertainty and computational feasibility.
The five frameworks that make up Bayesian econometrics today are not a single unified school but a collection of approaches that share core commitments while diverging on key questions. What they agree on is foundational: Bayes' theorem as the correct way to update beliefs, simulation as the universal computational strategy, and the importance of representing uncertainty probabilistically. What they disagree about is the nature and role of the prior. Prior-Based Bayesian Econometrics treats the prior as a substantive input that should reflect genuine economic knowledge; Bayesian Time Series and VAR Econometrics treats it primarily as a shrinkage device to regularize high-dimensional models; Hierarchical Bayesian Econometrics treats it as an estimable population distribution; and Bayesian Model Averaging treats the prior over models as a necessary but fragile ingredient whose influence must be carefully assessed. This disagreement is not a sign of weakness but a productive tension that maps onto different empirical tasks. In macroeconomics, the Minnesota prior and its descendants remain dominant for forecasting, while structural DSGE models use theory-based priors. In microeconometrics, hierarchical models are standard for handling heterogeneity, and BMA is increasingly used for variable selection in high-dimensional settings. The field's leading edge is the integration of these frameworks: hierarchical time-series models with shrinkage priors, BMA applied to hierarchical models, and simulation methods that make all of this computationally feasible. Bayesian econometrics has not resolved the question of what makes a good prior, but it has built a flexible toolkit in which that question can be asked precisely for each new problem.