Overview
Adrian E. Raftery is Professor of Statistics and Sociology at the University of Washington, Seattle. His research bridges Bayesian computation, model selection, and demographic forecasting. He developed or co-developed several widely used methods for assessing MCMC convergence (Raftery-Lewis 1992), computing Bayes factors from simulation output (Lewis-Raftery 1997), and Bayesian model averaging.
Key Contributions / Features
- Raftery-Lewis (1992): Diagnostic for determining how many Gibbs sampler iterations are required to estimate a quantile q of a target variable to within ±r with probability s. Computes required burn-in, thinning, and total iterations from a short pilot run; widely implemented in the
coda package. See Raftery-Lewis (1992).
- Kass-Raftery (1995): Comprehensive survey of Bayes factors — definition, computation methods (analytic, Laplace, importance sampling, MCMC), calibration via Jeffreys's (1961) qualitative scale, and sensitivity to priors. Standard reference for practitioners comparing models via Bayes factors. See Kass-Raftery (1995).
- Lewis-Raftery (1997): Laplace-Metropolis estimator for computing integrated likelihoods from MCMC output — basic LM for fixed-effects models (posterior mode via multivariate median + robust variance estimate) and compound LM for hierarchical models (inner Laplace over random effects, outer LM over fixed effects + hyperparameters). Batch-means uncertainty assessment. See Lewis-Raftery (1997).
- Bayesian model averaging: Methodology for accounting for model uncertainty by averaging predictions across models weighted by their posterior probabilities; formalized in Raftery-Madigan-Hoeting (1997) and related papers.
- Raftery (1995) (Sociological Methodology 25: 111–163): The applied case for Bayesian model selection — P-values fail in large samples, stepwise selection misleads, single-model inference ignores model uncertainty; BIC makes it practical. See Raftery (1995).
- Petrone-Raftery (1997): With Sonia Petrone, a caution that the Dirichlet process's discreteness materially distorts partition inference and Bayes factors in change-point/mixture models. See Dirichlet Process Mixture.
- Gneiting-Raftery (2007): With Tilmann Gneiting, the definitive review of proper scoring rules — characterization via convex functions/Bregman divergences, the CRPS and energy score for continuous/multivariate forecasts, the interval score, and the log-score = log-Bayes-factor identity. See Gneiting-Raftery (2007).
Related