Norman J Arnold School of Public Health-Introduction to Bayesian.pptVIP

CDC 2003 cancer conference CDC 2003 cancer conference Introduction to Bayesian Mapping Methods Andrew B. Lawson Arnold School of Public Health University of South Carolina South Carolina congenital abnormality deaths 1990 Mapping issues Relative risk estimation Disease Clustering Ecological analysis Relative risk estimation SMRs (standardized mortality /morbidity ratios Some notation For each region on the map: yi is the count of disease in the ith region ei is the expected count in the ith region θi is the relative risk in the ith region The SMR is just smri = yi/ ei This is just an estimate of θi SMR problems Notoriously unstable Small expected count can lead to large SMRs Zero counts aren’t differentiated The SMR is just the data! Smoothing for risk estimation Modern approaches to relative risk estimation rely on smoothing methods These methods often involve additonal assumptions or model components Here we will examine only one approach: Bayesian modeling Bayesian Modeling Some statistical ideas: Likelihood…….we usually assume that counts of disease have a Poisson distribution so that yi has a Poisson distribution with expected value ei θi We usually write this as yi ~Pois(ei θi) for short The counts have a Poisson likelihood Likelihood The counts have a joint probability of arising based on the likelihood L(y, θ) : L(y, θ) is the product of Poisson probabilities for each of the regions This tells us how likely the data are given the expected rates (ei θi) It also tells us what the most likely values of θ are given the data observed. Maximum Likelihood The SMR is the value of θ which gives the highest likelihood for the data (under a simple Poisson model)….this is called maximum likelihood (ML) This approach is often used in statistics to get good estimates of parameters Here we go beyond ML Smoothing using Bayesian methods One way to produce smoother relative risk estimators is to assume that the risk has a distribution In Bayesian terms this is c