Nonparametric bayesian multivariate meta-regression

Abstract

In biomedical research, meta-analysis is a popular tool to combine evidences from multiple studies to investigate an exposure-response association. A two-stage analytical approach is used in meta-analysis for its computational convenience and flexibility. The first stage estimates the association for each study while the second stage combines the study-specific estimates correcting for the study-specific error. The second stage often incorporates study-specific covariates (meta-predictors) and is called meta-regression. One application where the two-stage meta-analysis has been widely used is an epidemiological study for the health effects of environmental exposure, which often analyzes time-series data of exposure and health outcome collected from multiple locations. The first stage models location-specific association, which is often represented by multiple parameters as the association is nonlinear or delayed, and the second stage conducts a multivariate meta-regression with location-specific characteristics as meta-predictors. The currently used multivariate meta-regression is a form of multivariate normal linear regression, which may be limited as it assumes linearity in meta-predictors, residual normality and homoscedasticity. Hence, in this dissertation, we propose a flexible multivariate meta-regression in a nonparametric Bayesian modeling framework incorporating a residual spatial dependency. In chapter 2 and 3, meta-analysis and nonparametric Bayesian modeling were briefly reviewed. In chapter 4, two proposed models were presented in detail. In chapter 5, the proposed meta-regression models and currently used meta-regression model were applied to investigate a temperature-mortality association in the 135 US cities for comparison. In chapter 6, simulation study was conducted under three different scenarios for model evaluation.