Meta-Analysis Using Dirichlet Process
Office of Biostatistics
Center for Drug Evaluation & Research
U.S. Food and Drug Administration
Engineering Building, Room 1602
Time: 11 am – 12 pm
Date: Friday, April 20, 2012
A Bayesian approach for meta-analysis using the Dirichlet process is presented. The key aspect of the Dirichlet process (DP) in meta-analysis is the ability to capture the heterogeneity among studies while relaxing the distributional assumptions. We assume that the study effects are generated from a Dirichlet process. Under a DP model, the study effects parameters have support on a discrete space and enable borrowing information from study to study while facilitating clustering among studies. We evaluate the performance of the DP approach through simulations and illustrate the proposed method by applying it to three datasets; one large dataset on Program for International Student Assessment (PISA) involving 30 countries, a small dataset from published literature on the treatment of Alzheimer’s disease and a two-arm clinical trial dataset on preventing mortality after myocardial infarction. Results from the data analyses, simulation studies, and log-pseudo marginal likelihood (LPML) model selection procedure indicate that DP model perform better over conventional alternative methods.