Maximum Likelihood--Based Inference of Metadata via EM Algorithm

• Activity Number: 454
• Type: Topic Contributed
• Date/Time: Wednesday, August 4, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Ming-Hui Chen*+ and Joseph G. Ibrahim and Arvind K. Shah and Jianxin Lin
• Companies: University of Connecticut and The University of North Carolina at Chapel Hill and Merck Research Laboratories and Merck Research Laboratories
• Keywords: Aggregate covariates ; Meta normal regression models ; Meta variable selection ; Multiple trials ; Random effects

Abstract

The use of random effects models for meta-analysis has become a well accepted standard for analyzing meta data from multiple studies. However, the existing maximum likelihood based meta-analytic methods in the literature may not be sufficient to accommodate complex modeling schemes of random effects. When random effects are multidimensional, the existing algorithms for computing MLEs often converge poorly or even do not converge. We extend the method of Ibrahim, Chen, and Lipsitz (2001) to develop an efficient EM algorithm for fitting multidimensional random effects meta normal regression models. In addition, we develop a novel strategy for carrying out variable selection of aggregate covariates. A large meta data from clinical trials is used to illustrate the proposed methods. The proposed methodology can be extended to the meta-analytic models with missing aggregate covariates.