贝叶斯层次模型在嵌套结构调查数据中的应用研究  

Application of Hierarchical Bayesian Model for Nested Structural Epidemiological Data

作　　者：文雯[1] 文小焱 胡珊[1] 彭斌[1] 

机构地区：[1]重庆医科大学公共卫生与管理学院卫生统计学教研室

出　　处：《中国卫生统计》2015年第2期190-193,共4页Chinese Journal of Health Statistics

基　　金：国家自然科学基金(81373103);重庆市科委基础与前沿研究计划项目(cstc2013jcyj A10009)

摘　　要：目的针对分层抽样流行病调查数据的结构特点,构建两种基于分层嵌套思想的贝叶斯层次模型,并探讨其优缺点。方法以贝叶斯层次模型为基础,利用嵌套结构中的层级关系构建模型,其中,模型一以嵌套层效应分解为特点构建,模型二以嵌套层效应逐级传递为特点构建。以重庆市出生缺陷调查数据为例,采用Open BUGS软件进行模型拟合及分析。结果以偏差信息准则(deviance information criterion,DIC)作为拟合优度评价,模型一和模型二的DIC值分别为101.8和101.6,大致相等;敏感性分析显示,在总体率的超参数μ设置不同先验信息下,模型一和模型二对总效应估计的变异性分别为(用标准差度量,10-4):后验均数1.191和27.546;后验中位数1.038和7.617,模型一的变异性比模型二小。结论模型一和模型二均可用于嵌套结构的调查数据建模分析及预测,拟合效果相当;但模型一比模型二受先验信息影响小,稳健性更好,更适合先验信息欠缺时的数据分析。Objective To develop two hierarchical Bayesian models for the epidemiological data with focus ing on its nested structure; as well as to explore the pros and cons of them. Methods Relationships among nested layers of nested structu ral data are taken into account when developing the two hierarchical Bayesian model s. The first model focuses on the stratification effect of each nested layer for differentiation between the layers. The second model focuses on the transmission effect between the father layer and its son layers. OpenBUGS software and a b irth defects s urvey data were used to fit and evaluate the two hierarchical Bayesian models; and the d eviance information criterion （DIC） was used for measuring the goodness-of-fit of them . A sensitivity analysis was conducted with different sets of prior information on hyper parameter of the population rate μ.Results The DIC of the two models are 101.8 and 101.6, respectively, which shows almost the same goodness-of -fit of them. The sensitivity analysis shows that the standard deviation of the two models for t he posterior mean of estimated population rate θ0are（10-4）1.191and 27.546, respectively, for the posterior median of them are（10^-4）1.038and7.617, respectively. Both results of posterior mean and posterior median say that the first model has smaller standard deviation under different prior information scenario. Conclusion Both models can be used to model nested structural epidemiological data. However, the first model is affected by prior information much less than the second model does . Thus, the first model is more stable and is better to model nested structural survey data when little prior information is available .

关 键 词：嵌套结构数据 分层抽样 贝叶斯层次模型 OpenBUGS 

分 类 号：O212.8[理学—概率论与数理统计] R195.1[理学—数学]