Bayesian Econometrics

Bayesian econometrics is the branch of econometrics that treats every unknown quantity—a coefficient, a variance, a policy parameter, even the choice of model—as a random variable described by a probability distribution.

11 frameworks, from 1946 to today · Econometrics · Economics

still a working view ended · each leads to its entry below

Overview

Bayesian econometrics is the branch of econometrics that treats every unknown quantity—a coefficient, a variance, a policy parameter, even the choice of model—as a random variable described by a probability distribution1. Where classical (frequentist) econometrics asks how often a procedure would be right across repeated samples, the Bayesian asks what the data, combined with prior knowledge, imply about the quantities of interest12. The answer is a posterior distribution: a probability statement about the parameters given the data actually observed1. What is at stake is how economists reason under uncertainty—whether prior knowledge is a legitimate input, how to forecast, and how to compare rival theories. The history of the field is largely a history of overcoming a computational obstacle: for most of the twentieth century the mathematics of Bayes' rule was simple but the integrals it required were impossible to evaluate, until simulation methods changed everything23.

Bayes' Rule and the Early Case for Subjective Probability

The foundation is Bayes' theorem, published posthumously by Thomas Bayes in 1763 and generalized by Pierre-Simon Laplace, who used it to estimate unknown quantities in astronomy and demography. The rule is a piece of arithmetic: the posterior is proportional to the likelihood (how probable the observed data are under each parameter value) times the prior (what was believed before seeing the data)1. Laplace applied it widely, but the method fell out of favor in the nineteenth and early twentieth centuries as statisticians sought "objective" procedures that did not require a prior.

The modern revival in economics came through decision theory. Leonard Savage's The Foundations of Statistics (1954) gave an axiomatic justification: an agent whose choices obey certain consistency conditions behaves as if she holds a subjective probability distribution and maximizes expected utility4. This made subjective probability respectable for economists, who were already comfortable with expected-utility reasoning. In econometrics the case was pressed by Arnold Zellner, whose textbook An Introduction to Bayesian Inference in Econometrics (1971) became the field's founding reference. Zellner argued that the Bayesian framework handled nuisance parameters, sequential learning, and model uncertainty more coherently than sampling theory, and he built much of the early toolkit himself.

Conjugate Priors and the Age of Closed Forms

Because the required integrals were intractable, early Bayesian econometrics depended on conjugacy: choosing a prior distribution of a mathematical form that combines with the likelihood to yield a posterior of the same family, so the answer can be written down in closed form. The workhorse is the normal–linear model, where a normal prior on regression coefficients paired with a known or inverse-gamma-distributed variance yields a normal posterior. Zellner's work on this "natural conjugate" setup, and the parallel development of the normal-gamma and inverted-Wishart families, let economists compute posteriors by hand or with a desk calculator.

The limitation was severe: conjugate analysis worked only for a narrow class of models, and the priors it required were often convenient rather than defensible. This tension drove the next debate.

The Battle over Priors: Objective, Reference, and Hierarchical

If the prior is arbitrary, the posterior is arbitrary—so a large strand of the field sought priors that let the data "speak." Harold Jeffreys in the 1930s and 1940s proposed improper and invariant priors (rules that give the same answer regardless of how a variable is measured), and in econometrics Zellner and others developed "diffuse" or "noninformative" priors intended to represent ignorance. The difficulty is that ignorance is hard to formalize: a flat prior on one parameter implies a non-flat prior on its transformations, and improper priors can produce improper (non-integrable) posteriors.

José Bernardo (1979) proposed reference priors, designed to maximize the information the data contribute, and James Berger and colleagues developed the theory further through the 1980s and 1990s. A different resolution was hierarchical modeling: instead of fixing a prior, treat its parameters (hyperparameters) as themselves drawn from a higher-level distribution, so the data inform how much prior weight to give5. This multilevel structure, formalized in the 1970s and widely adopted in the social sciences, let economists pool information across units—states, firms, countries—borrowing strength where data are thin. The debate between subjective, objective, and hierarchical priors never fully closed; it remains a live methodological choice.

The Minnesota Prior and Bayesian Vector Autoregressions

A decisive practical success came in macroeconomics. Christopher Sims's vector autoregression (VAR), introduced in 1980, modeled each variable as a function of its own and others' past values—flexible but parameter-heavy, so that unrestricted VARs overfit and forecast poorly. Robert Litterman, working at the Federal Reserve Bank of Minneapolis, proposed in the early 1980s a Minnesota prior: shrink coefficients toward the belief that each variable follows a random walk (its own recent value, with others' lags contributing little), with the strength of shrinkage tuned by a small number of hyperparameters. The result, the Bayesian VAR, forecast better than unrestricted VARs and became a standard tool at central banks. It was an early demonstration that a shrinkage prior—one that pulls estimates toward a simple benchmark—could beat both classical and naive alternatives, an idea that later flowered into regularization across econometrics.

Simulation Breaks the Computational Barrier

The real transformation came from computing1. In the late 1980s and early 1990s, statisticians developed Markov chain Monte Carlo (MCMC): instead of computing the posterior integral directly, draw a long sequence of random values whose long-run distribution is the posterior, then summarize it by averaging. The Metropolis–Hastings algorithm (1953) and the Gibbs sampler (popularized by Geman and Geman in 1984, and brought into statistics by Gelfand and Smith in 1990) made this practical. The Gibbs sampler is especially suited to econometrics because many models decompose into conditional distributions that are conjugate even when the joint posterior is not.

The effect was to dissolve the closed-form constraint5. Nicholas Polson, in a 1990s body of work with various coauthors, showed how to fit models with regime switching, stochastic volatility, and heavy-tailed errors by augmenting the parameter space with latent variables and sampling them. The Bayesian approach now handled models that frequentist methods could barely estimate. A parallel line, Monte Carlo integration more generally—using random draws to approximate integrals—underpinned all of it.

Model Comparison, Averaging, and Diagnostics

Once models could be estimated, the question of which model arose. The Bayesian answer is the posterior odds: the ratio of each model's marginal likelihood (the probability of the data under the model, averaging over its parameters)67. This is elegant but sensitive to priors, and computing marginal likelihoods is hard—leading to approximations such as the Bayesian information criterion and, later, bridge sampling and nested sampling. Bayesian model averaging takes the logic further: rather than pick one model, weight all candidate models by their posterior probability and average their predictions8. Developed in the 1990s (with influential work by Adrian Raftery and colleagues), it addressed the "which regressors?" problem in growth economics and forecasting, and it made explicit that model uncertainty is itself a quantity to be reported. Diagnostics—checking whether a fitted model reproduces features of the data—grew alongside, using posterior predictive checks and convergence measures for MCMC chains.

Bayesian Macroeconomics: DSGE Estimation

The most ambitious application came in macroeconomics. Dynamic stochastic general equilibrium (DSGE) models—microfounded models of the whole economy—were traditionally calibrated rather than estimated, because their likelihoods were intractable. In the 2000s, work by Frank Schorfheide, Jesús Fernández-Villaverde, Juan Rubio-Ramírez, and others showed how to estimate them by Bayesian methods: solve the model, linearize it into a state-space form, evaluate the likelihood with the Kalman filter, and sample the posterior with MCMC. This let macroeconomists bring formal theory and data together, compare competing model specifications, and produce forecasts with error bands. Bayesian DSGE estimation became a standard tool at central banks and in academic macro, and it remains a benchmark against which newer methods are judged.

Nonparametrics and the Modern Toolkit

Two further developments broadened the field. Bayesian nonparametrics replaces fixed finite-dimensional parameters with infinite-dimensional objects—random functions, Dirichlet processes, infinite mixtures—so that model complexity grows with the data. The Dirichlet process, introduced by Thomas Ferguson in 1973, and later developments such as the hierarchical Dirichlet process and Gaussian-process priors, gave economists flexible tools for density estimation, clustering, and heterogeneous-agent modeling without committing to a fixed number of components.

Most recently, variational inference and probabilistic programming have changed how Bayesian models are built and fit9. Variational methods replace sampling with optimization: approximate the posterior by the closest member of a tractable family, trading exactness for speed on large datasets10. Probabilistic programming languages (Stan, PyMC, and others) let a researcher write a model declaratively and have the software choose an inference algorithm91112. These tools have lowered the barrier to entry and pushed Bayesian methods toward the scale of modern economic data.

The Present

Bayesian econometrics today is less a separate school than a set of methods woven through the discipline. MCMC and its descendants remain the default for structural macro and time-series work; hierarchical models are standard in panel and multilevel settings; Bayesian model averaging and shrinkage priors are routine in forecasting and in high-dimensional regression; and variational and probabilistic-programming approaches are spreading as datasets grow. The old debates persist in new form—how much prior information is legitimate, how to validate a model that was fit by simulation, how to scale inference to very large problems—and the field's history is, in large part, the record of successive answers to the question of what a probability statement about an unknown parameter should mean.

References

  1. Bayesian econometrics (introduction, Basics, History), Wikipedia
  2. Bayesian statistics, Wikipedia
  3. Bayesian inference (History), Wikipedia
  4. Subjective expected utility, Wikipedia
  5. Prior probability, Wikipedia
  6. Bayes factor, Wikipedia
  7. Marginal likelihood, Wikipedia
  8. Ensemble learning (Bayesian model averaging), Wikipedia
  9. Probabilistic programming, Wikipedia
  10. Variational Bayesian methods, Wikipedia
  11. Stan (software), Wikipedia
  12. PyMC, Wikipedia

The frameworks of Bayesian Econometrics

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1946-Present

    Objective and Reference Priors

    What prior lets the data speak for themselves?

    Fly to its tower in Nousfarer

  2. 1954-Present

    Subjective Bayesian Decision Theory

    How should beliefs be represented and updated?

    Fly to its tower in Nousfarer

  3. 1960-1990

    Conjugate prior

    When can a posterior be written in closed form?

    Fly to its tower in Nousfarer

  4. 1972-Present

    Bayesian hierarchical modeling

    How can units borrow strength from each other?

    Fly to its tower in Nousfarer

  5. 1973-Present

    Bayesian Nonparametrics

    How can models avoid fixed finite parameterizations?

    Fly to its tower in Nousfarer

  6. 1978-1995
  7. 1980-Present

    Shrinkage Priors and Bayesian VARs

    How can many coefficients be shrunk toward a benchmark?

    Fly to its tower in Nousfarer

  8. 1990-Present
  9. 1990-Present

    Markov Chain Monte Carlo

    How can we sample from an intractable posterior?

    Fly to its tower in Nousfarer

  10. 2000-Present

    Bayesian DSGE Estimation

    How can formal macroeconomic theory be fitted to data?

    Fly to its tower in Nousfarer

  11. 2010-Present

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Objective and Reference PriorsSubjective Bayesian Decision TheoryConjugate priorBayesian hierarchical modelingBayesian NonparametricsMonte Carlo integration for Bayesian …Shrinkage Priors and Bayesian VARsBayesian Model Comparison and Averagi…Markov Chain Monte CarloBayesian DSGE EstimationVariational Inference and Probabilist…

Beyond the field

Where Bayesian Econometrics's ideas came from in other fields, and what they led to there: the strongest 12 of 36 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Bayesian Econometrics is a planet in Nousfarer

Its city has a tower for each of these 11 frameworks. Land, walk in, and climb one concept at a time.

Land on Bayesian Econometrics

Related fields

Fields its ideas are linked to

More in Econometrics

See where Bayesian Econometrics sits on the Map of Human Knowledge