John Geweke

Overview

John Geweke is a Bayesian econometrician known for developing posterior simulation methods and applying them to structural time-series and panel data models. He held positions at the University of Iowa, University of Minnesota, and later the University of Technology Sydney. His work encompasses Monte Carlo integration, spectral convergence diagnostics for MCMC, and flexible non-parametric Bayesian approaches to distributional modeling.

Key Contributions

Geweke (1989): Bayesian inference in econometric models via Monte Carlo integration; foundational reference for posterior simulation in econometrics.
Geweke (1992): Spectral variance estimator for MCMC convergence diagnostics — $V(\hat\mu_i) = S(0)/k$ where $S(0)$ is the spectral density of the chain at frequency zero. Referenced in Gibbs Sampler.
Geweke (1993): Exact Bayesian inference for the Student-t linear model via the scale-mixture-of-normals representation ( $\nu/\omega_i \sim \chi^2(\nu)$ ); four-block Gibbs sampler with conjugate conditionals; convergence Theorems 4–5; posterior odds 13/14 Nelson-Plosser series favor Student-t; unit root conclusions shown to be sensitive to tail assumptions. See Geweke (1993).
Geweke (1996): Bayesian inference for the normal linear model with linear inequality constraints $a \leq D\beta \leq w$ ; two algorithms — GHK probability simulator for evaluating constraint probability $p_{2|1}$ (always more efficient than accept-reject; 265× faster for tight constraints); Gibbs sampler via coordinate-wise truncated normal draws after reparameterizing $z = D(\beta-b)$ . Extensions to scale-mixture models, SUR, and multinomial probit. See Geweke (1996).
Geweke-Keane (1999): Mixture of normals probit model — replaces the $\mathcal{N}(0,1)$ probit disturbance with a finite $m$ -component Gaussian mixture; six-block conjugate Gibbs sampler with dual augmentation (latent utilities $\tilde{y}_t$ from Albert-Chib 1993b and component indicators $L_t$ ); Gelfand-Dey marginal likelihood on the observed-data parameter space; PSID female LFP application: BF $\approx 10^7$ against conventional probit, best model is scale mixture of 4 normals. See Geweke-Keane (1999).
Geweke (1999): Comprehensive treatment of simulation-based Bayesian inference; introduced the relative numerical efficiency (RNE) concept for comparing MCMC samplers.
Geweke-Keane (2000): Dynamic panel earnings model (PSID 1968–1989) with 3-component mixture of normals for shocks; Gibbs sampler with data augmentation; mixture changes lifetime earnings gaps and poverty persistence substantially vs. normal model.
Geweke-Keane (2007): Smoothly Mixing Regressions — finite mixture of normal regressions where state probabilities depend on covariates via multinomial probit; five-block Gibbs with Metropolis-within-Gibbs for latent weight matrix; single-dimension quadrature for state probability integrals; full conditional distribution recovers posterior quantile surfaces; PSID earnings and S&P 500 applications (SMR beats t-GARCH by 57 log-points out-of-sample). See Geweke-Keane (2007) and Smoothly Mixing Regressions.
Geweke-Gowrisankaran-Town (2003): Bayesian simultaneous-equation selection model for hospital quality; multinomial probit for hospital choice (instrumented by patient-hospital distance) and binary probit for 10-day in-hospital mortality; errors linked via hospital-specific severity correlations $\delta_j$ ; Gibbs sampler with data augmentation for 74,848 Medicare patients and 114 Los Angeles County hospitals; selection component of mortality variance 8× larger than independent component; quality U-shaped in bed size; standard probit rankings severely misleading. See Geweke-Gowrisankaran-Town (2003) and Selection Model.

John Geweke

Overview

Key Contributions

Related