"Introduction to Bayesian Hierarchical and Multi-level Models" extends the Bayesian modeling framework to cover hierarchical models and to add flexibility to standard Bayesian modeling problems. Participants will learn how to define three stage hierarchical models and to implement them using Winbugs, in multilevel, meta-analytic and regression applications. Continuous, count and binary outcomes are covered. Participants will also learn how to assess goodness-of-fit.
Dr. Peter Congdon is a Research Professor in Quantitative Geography and Health Statistics at Queen Mary University of London. He is the author of several books and numerous articles in peer-reviewed journals. His research interests include spatial data analysis, Bayesian statistics, latent variable models and epidemiology.
- Overview of application contexts: meta-analysis to summarise accumulated evidence; comparisons of related units (e.g. "league table comparisons" of exam results, hospital mortality rates, etc); rationale for multi-level models in health, education etc
- Defining Hierarchical Bayesian Models. Three stage models.
- Benefits from "borrowing strength" using Bayesian random effect models.
- Measuring model fit for hierarchical models, and procedures for model checking; effective parameters (and DIC)
- Common conjugate hierarchical models with worked examples
- Modelling the variance/covariance in Bayesian random effects models. Alternative priors for variances. Winbugs implementation of these priors.
- Bayesian meta-analysis and pooled estimates in clinical studies and education
- Different meta-analysis schemes (e.g. beta-binomial, logit-normal for binomial data)
SESSION 3 - Multi-Level and Panel Models
- Multi-level models (2 and 3 level models for continuous, count and binary responses) and Winbugs implementation to include data input structures.
- Simple panel models (random intercept, random slope) from a Bayesian perspective.
- Crossed and multivariate and multilevel models
- Overdispersed regression options for count and proportion data including negative binomial and beta-binomial regression