Bayesian inference for multivariate meta-regression with a partially observed within-study sample covariance matrix

摘要

The article studies the problem related to estimating the within-study covariance matrix in multivariate meta-regression. Reviewing some of the disadvantages and difficulties prevailing in existing approaches, a novel methodology based on the Bayesian approach for estimating the within-study covariance matrix in multivariate meta-regression is proposed. As an example, the cholesterol meta-data with three primary aggregate outcome variables from the 26 clinical trials is considered followed by which the multivariate meta- regression random effects model is developed. A novel Markov Chain Monte Carlo sampling algorithm is developed based on sampling of the partial correlations. A specific type of structures and models for the unknown within- Study covariance matrix that may be used in practice is studied. The general Bayesian computational development and goodness-of-fit criterion for model comparisons is presented. Two simulation studies are conducted to perform analysis of the proposed method. The results from both the simulation studies and the cholesterol meta-data example are discussed in detail.