CMStatistics 2016: Start Registration
View Submission - CMStatistics
B0903
Title: Nonparametric Bayesian multivariate meta-regression: An application in the temperature-mortality association study Authors:  Gyuseok Sim - Korea Advanced Institute of Science and Technology (Korea, South) [presenting]
Yeonseung Chung - Korea Advanced Institute of Science and Technology (Korea, South)
Ho Kim - Seoul National University (Korea, South)
Antonella Zanobetti - Harvard School of Public Health (United States)
Joel Schwartz - Harvard School of Public Health (United States)
Abstract: In biomedical research, meta-analysis has been a popular and useful tool combining evidence from multiple studies to investigate an exposure-response association. Because of the hierarchical nature in meta-analysis, a two-stage analytical approach has been used in many studies for the computational convenience. In the first stage, study-specific exposure-response relationships are estimated as multi-parameters using individual data. These estimates are combined through meta-analysis in the second stage incorporating study-level predictors (called meta-predictors). The second stage is often called meta-regression. The currently used multivariate meta-regression assumes linearity in meta-predictors, residual normality and heteroscedasticity. However, such assumption is limited to integrate the study-specific estimates over a broad range of studies where sub-groups may exist and the effects of meta-predictors are not linear. We propose a more flexible and generalized two-stage multivariate meta-regression by replacing the second stage by a nonparametric Bayesian multivariate nonlinear regression based on the Dirichlet process mixture. The proposed method was evaluated through a simulation study and applied to the data from 135 US cities to study the temperature-mortality association.