Edward I. George

Overview

Edward I. George is a statistician at the University of Pennsylvania (Wharton School), formerly at the University of Texas at Austin. He is known for Bayesian variable selection methods, for co-authoring the canonical tutorial on the Gibbs sampler, and for foundational work on conjugate likelihood distributions for hierarchical model hyperparameter sampling.

Key Contributions

• "Explaining the Gibbs Sampler" (1992) (with George Casella): Tutorial derivation of Gibbs convergence; fixed-point equation for conditionals → marginals; Rao-Blackwell density estimator; failure mode analysis.
• George-McCulloch (1993): Stochastic Search Variable Selection (SSVS) — three-block Gibbs over $(\beta, \sigma^2, \gamma)$ with mixture prior $\beta_i|\gamma_i \sim (1-\gamma_i)\mathcal{N}(0,\tau_i^2)+\gamma_i\mathcal{N}(0,c_i^2\tau_i^2)$. The $\gamma_i$ block is Bernoulli with probability $a/(a+b)$ where $a,b$ evaluate the spike/slab Normal density at the current $\beta^j$ draw (no direct dependence on $Y$). All post-burn-in $\gamma$ patterns identify high-probability subsets across the $2^p$ model space without enumeration. Demonstrated for $p=60$ regressors.
• George-McCulloch (1997): Unified hierarchical framework embedding SSVS; conjugate prior $\pi(\beta|\sigma,\gamma)=\mathcal{N}_p(0,\sigma^2 D^*_\gamma R_\gamma D^*_\gamma)$ (eq. 15) links $\beta$ and $\sigma$, enabling analytical integration over $(\beta,\sigma^2)$ and exact $g(\gamma)$ formula (eq. 24). Chambers (1971) and Dongarra (1979) fast updates ($O(p^2)$ per $\gamma_i$ flip) enable Gray Code exhaustive evaluation for $p \leq 25$ ($p=20$: 5 min, 1996 hardware). Three-algorithm comparison (NG/CG/CM); stuck-$\gamma$ failure mode of NG under weak information (MC SE ≈ 0.10 vs. CG/CM ≈ 0.02); normalization constant estimation; two-phase S&P 500 application ($p=200$, 21-variable best model, $R^2=95.6\%$, stepwise model posterior = $0.00176\times$ maximum). Statistica Sinica 7 (1997): 339–373.
• George-Makov-Smith (1993): Established the theory of conjugate likelihood distributions for exponential-family hierarchical models. Theorem 3.3: the hyperparameter full conditional $\pi(\alpha,\beta|\theta_1,\ldots,\theta_p)\propto L(\alpha,\beta|\theta_1,\ldots,\theta_p)\cdot\pi(\alpha,\beta)$ is log-concave (enabling Gilks-Wild ARS) and jointly proper iff $p\geq 2$. Corollary 3.5 gives the explicit Gamma-Poisson formula; pump-failure application (p=10) illustrates the complete exact Gibbs cycle.
• George-Sun-Ni (2008): Extended SSVS to VAR models, simultaneously searching restrictions on the coefficient matrix Φ and the off-diagonal elements of the upper-triangular Cholesky factor Ψ (where Σ⁻¹ = ΨΨ'). Upper-triangular Ψ ensures global identification. Five-step Gibbs sampler with all conditionals in standard form (Gamma, Normal, Bernoulli — no MH steps). Key property from George-McCulloch (1993) preserved: $\gamma$ indicators for Φ do not depend on $Y$. Rao-Blackwell model-averaged forecasts improve over MLE by 20–67% in simulation; empirical application to 7-variable PPI→CPI VAR identifies contemporaneous supply-chain price transmission structure (1969–80) and structural change post-1981 (Volcker disinflation). Journal of Econometrics 142 (2008): 553–580.

Related

• Gibbs Sampler
• Variable Selection
• Bayesian Hierarchical Model
• Vector Autoregression
• George Casella
• Robert E. McCulloch
• Udi E. Makov
• A.F.M. Smith
• Dongchu Sun
• Shawn Ni
• George-McCulloch (1993)
• George-McCulloch (1997)
• George-Makov-Smith (1993)
• George-Sun-Ni (2008)