Overview
Bradley P. Carlin is a Bayesian statistician, formerly at the University of Minnesota School of Public Health and later at the Mayo Clinic. He is known for foundational work in Bayesian computation and biostatistics. In econometrics his primary contribution is the Carlin-Chib (1995) pseudo-prior algorithm for Bayesian model choice via MCMC, co-authored with Siddhartha Chib.
Key Contributions
- Chib-Carlin (1999) blocked Gibbs for longitudinal models — with Siddhartha Chib; proposes partially and fully blocked MCMC algorithms for the Gaussian linear mixed model and binary longitudinal probit; Algorithm 2 marginalizes fixed effects β over random effects for essentially iid draws; Algorithm 3 evaluates the integrated likelihood via the Chib (1995) identity as a ratio of three Normal densities and updates variance components jointly by M-H; Algorithm 7 fully marginalizes over random effects and latent binary data via the GKH method; three empirical examples (HIV CD4, Six Cities wheezing, PSID labor-force). See Gibbs Sampler and Bayesian Hierarchical Model.
- Carlin-Chib (1995) model-augmented Gibbs sampler — pseudo-prior approach for Bayesian model comparison: introduces linking densities for inactive parameters so a single Gibbs chain visits all K candidate models; estimates posterior model probabilities directly from visit frequencies; Bayes factors follow as visit frequency ratios. See Gibbs Sampler and Marginal Data Density.
- Han-Carlin (2001) comparative review — with Cong Han, benchmarked six MCMC methods for Bayes factor computation across three empirical examples (radiata pine, AIDS CD4 counts, prostate cancer); documented the breakdown of product-space samplers (CC/MCC) for models with many random effects (467 subjects, 1.5M iterations, no convergence); recommended marginal likelihood identity methods for standard hierarchical models and reversible jump for variable-dimension problems.
Related