环境流行病学中多元非线性高维交互效应的贝叶斯模型构建及推断研究

Outdoor air pollution in China has aroused wide health concerns, and air pollution is a complex mixture of gaseous, liquid, and solid components that vary greatly in composition and concentration across the China and around the world due to differences in sources, weather, and topography. The challenges of determining whether effects are additive, synergistic, or less-than-additive, and of identifying climatic effect modification in epidemiologic studies, are substantial. Often, a high degree of correlation exists among levels of different pollutants emitted from similar sources or generated through similar atmospheric processes; and there may be nonlinear interactions among pollutants in relation to health outcomes. These issues complicate and may even preclude the use of classic regression approaches. .The aim of this study is to develop innovative statistical methods for studying the combined effects of individual pollutants under different climate condition. Specifically, this project will develop a family of Bayesian hierarchical models to simultaneously account for multivariate pollutants, effect modification of climatic variables. We first construct the stage 1 likelihood model using semiparametric Gaussian Process and high dimensional tensor product, which will be the basis for the stage 2 prior model. In the prior model, multi-stage Gaussian covariance function and Spike-and-Slab prior will be used for effective regularization of the multivariate exposure–response surface. Next, we developed an extension of hybrid Gibbs sampling approach that incorporates two MCMC optimization strategies, adaptive sampling and latent variable augmentation. Then, a class of Bayesian predictive model checking methods will be established. Finally, health effect of single pollutant under complex model will be constructed under the framework of average predictive comparisons..All methods proposed were required to include validation of the approach either by using simulation studies or by conducting a thorough sensitivity analysis with widely available data sets. The proposed statistical methods can facilitate the understanding of the health impact of multivariate air pollutants under variable climate condition, thereby informing policy making.

定量阐明我国大气污染与人群健康效应的暴露-反应关系规律是大气污染健康风险评估和环境空气质量标准制修订的重要科学依据。公众健康受到多样气象条件及多种大气污染物的影响，而目前统计方法的不足限制了不同气象条件下多元大气污染物健康效应的精确估计。.本项目拟提出并构建一套新的层次贝叶斯统计模型及其推断方法：①先验分布引入多阶段高斯过程协方差函数与高维数据贝叶斯变量选择两种方法，并分别运用多种先验分布结构以探讨各方法的适用性；②提出多种基于后验预测分布的模型诊断统计量及贝叶斯P值计算方法并获得最优检验策略；③建立复杂交互非线性模型下单一污染物的健康效应新指标评价方法。后验分布运用Hybrid Gibbs抽样结合自适应规则与潜变量设计两种改进优化算法。项目将揭示多元大气污染在不同气象条件下健康影响的暴露-反应规律，具有重要科学意义；还为我国建立大气污染暴露基准建议值及相关政策提供技术支持，具有实用价值。

定量阐明我国大气污染与人群健康效应的暴露-反应关系规律是大气污染健康风险评估和环境空气质量标准制修订的重要科学依据。公众健康受到多样气象条件及多种大气污染物的影响，而目前统计方法的不足限制了不同气象条件下多元大气污染物健康效应的精确估计。.本项目以四川省作为研究现场，呼吸系统死亡与心血管疾病死亡作为健康结局代表，生态学时间序列数据设计为例进行统计方法学研究。研究对大气污染物健康效应关系的关键生物学参数及其共变规律的数据特征进行了探讨，并基于滞后机制、层次贝叶斯模型刻画了大气污染物与气象因子和健康效应之间的时空关联，且以分步计算策略为基础对贝叶斯算法进行优化，从而加快了层次模型中关键流行病学参数的后验分布的计算速度。研究从暴露尺度估计、综合指标建立和结果报告指标出发，采用层次贝叶斯空间模型、贝叶斯加权模型等对空气污染物浓度进行了更准确的估计，并建立了新型暴露估计指标和AQHI指标。研究还对不同联合效应结构下的平均预测差异指标的不同权重函数的应用价值进行了探讨，对基于后验预测分布诊断理论的模型诊断研究了多种统计量组合诊断方式以期提高识别模型与数据间差异的功效。.本项目从关键流行病学参数的设定、非线性效应的最大形式、各种变量滞后效应时间范围，到贝叶斯模型的构建、后验分布的高效推断算法，再到基于变量后验预测分布的模型诊断、单一污染物的平均健康效应评价函数的构建，直至各种模型算法的适用场景的实证与模拟研究，所有部分步步衔接，构成统一的有机整体，为类似的研究提供了新的思路，也为大气污染的防控提供了新策略。