Li Dalin, Conti David V
Department of Preventive Medicine and Zilkha Neurogenetic Institute, Keck School of Medicine, University of Southern California, Los Angeles, California 90089, USA
Am J Epidemiol. 2009 Feb 15;169(4):497-504. doi: 10.1093/aje/kwn339. Epub 2008 Dec 13.
The conventional method of detecting gene-environment interactions, the case-control analysis, suffers from low statistical power. In contrast, the case-only analysis/design can be powerful in certain scenarios, although violation of the assumption of independence between the genetic and environmental factors can greatly bias the results. As an alternative, Bayes model averaging may be used to combine the case-control and case-only analyses. This approach first frames the case-control and case-only analyses as variations of a log-linear model. The weighting between these 2 models is then a function of the data and prior beliefs on the independence of the 2 potentially interacting factors. In this paper, the authors demonstrate via simulations that when there is no prior information on the independence of the genetic and environmental factors, this approach tends to be more powerful than the case-control analysis. Additionally, when the genetic and environmental factors are not independent in the population, bias is substantially reduced, with a corresponding reduction in type I error in comparison with the case-only analysis. Increased power or increased robustness to violations of the independence assumption may be obtained with more appropriate prior specification. The authors use an example data analysis to demonstrate the advantages of this approach.
检测基因 - 环境相互作用的传统方法——病例对照分析，存在统计效能低的问题。相比之下，病例单组分析/设计在某些情况下可能很有效，尽管违反基因与环境因素之间独立性的假设会极大地使结果产生偏差。作为一种替代方法，贝叶斯模型平均法可用于结合病例对照分析和病例单组分析。这种方法首先将病例对照分析和病例单组分析构建为对数线性模型的变体。然后，这两个模型之间的加权是数据以及关于两个潜在相互作用因素独立性的先验信念的函数。在本文中，作者通过模拟表明，当没有关于基因与环境因素独立性的先验信息时，这种方法往往比病例对照分析更具效能。此外，当基因与环境因素在总体中不独立时，与病例单组分析相比，偏差会大幅降低，同时I类错误也相应减少。通过更合适的先验设定，可提高效能或增强对独立性假设违反情况的稳健性。作者通过一个示例数据分析来展示这种方法的优势。
