ベイズ流メタアナリシスにおける外れ値試験の検出と影響力解析

抄録

In evidence-based medicine, meta-analysis is a relevant method for research syntheses. Random-effects model has been a primary statistical tool for meta-analysis since it enables a quantitative evaluation of the treatment effect accounting for the between-studies heterogeneity. In practices of meta-analyses, some studies may have markedly different characteristics from the others and such “outlying” studies might yield misleading results. For this problem, although several frequentists’ methods to detect outlying studies have been developed, there has been no effective Bayesian method to detect outlying studies and to assess their influence. In this article, we proposed influence diagnostic methods for meta-analyses using four Carlin-Louis-type influence measures; (a) relative distance, (b) standardized residual, (c) Bayesian p-value, and (d) scale parameters in scale mixture models. We also demonstrated the practical effectiveness of these proposed methods through applications to four meta-analyses for a spinal manipulative therapy, renin angiotensin system inhibitors, a known history of gestational diabetes, and antenatal corticosteroids.