复杂抽样条件下健康调查数据的倾向评分匹配法研究

Complex survey is one of the most important study methods in health area. In complex survey, the selection bias is large, and the sampling weights are not equal, and the samples in the same cluster are always not independent, and there are a lot of confounders. For these reasons, traditional methods cannot control the confounding very well. Propensity score method can translate many confounders as a single score and control multiple confoundings. However, the effects of sampling weight and cluster cannot be comprehensively considered in the current propensity score theory. Theoretically, the current propensity score method cannot be applied in the analysis of complex survey data. Currently, few preliminary studies have just explored the method of incorporating sampling weight into propensity score or just discussed cluster effects under multilevel data framework. These researches have not comprehensively investigated the data structure of complex survey. On the basis of Monte Carlo simulation and real survey data of health status, this study will discuss quantitative methods to incorporate sampling weight and cluster effects into propensity score method both at score estimation stage and the score application stage, under complex survey framework. A new propensity score estimation model and a new propensity score matching method for complex survey data will be proposed. The methods we developed in this study can greatly improve the balance of the covariates between comparison groups and improve the precision of estimation of between-group effects, and will provide important methodology and practical reference for the application of propensity score matching method in complex health survey data.

复杂抽样调查是健康领域重要的研究方法之一，但是具有选择偏倚大、混杂因素多、抽样权重不等以及群内样本不独立的特征，传统的混杂和偏倚控制方法效果不佳。倾向评分法可将多个混杂因素综合为一个评分，从而控制多种混杂和偏倚。然而，现有的倾向评分法没有考虑抽样权重和群效应的影响，理论上还不能分析复杂抽样调查数据。目前的研究只单独地探讨抽样权重或者在多水平数据框架下探讨群效应，不能全面地反映复杂抽样调查数据的结构特征。本课题基于Monte Carlo模拟和城乡居民健康状况的真实调查数据，将抽样权重和群效应纳入同一个框架，探讨倾向评分的估计和应用过程中抽样权重和群效应的量化方法，提出复杂抽样条件下倾向评分的估计模型和匹配策略。该方法有助于提高复杂抽样数据协变量的组间均衡性和组间效应估计的准确性，将为倾向评分匹配法在复杂抽样健康调查研究中的应用提供重要的方法学依据和实践参考。

背景：复杂抽样调查具有选择偏倚大、混杂因素多、抽样权重不等以及群内样本不独立的特征，传统的倾向评分法没有考虑抽样权重和群效应的影响，理论上还不能分析复杂抽样调查数据。研究内容：本课题基于城乡居民健康状况的真实调查数据，将抽样权重和群效应纳入.同一个框架，探讨复杂抽样条件下倾向评分的估计模型和匹配策略。（1）比较加权复杂抽样校正法（CSw）、分层加权复杂抽样校正法（CSsw）、分层整群加权（CSscw）三种复杂抽样校正方法的抽样效率，分析不等概率、分层和整群3个复杂抽样特征对抽样效率影响大小；（2）开展了传统倾向评分匹配方法在两个大规模异质性人群慢性病患病率（高血压和糖尿病）比较中的应用研究；（3）基于陕西省第五次国家卫生服务调查数据，忽略复杂抽样特征采用传统倾向评分方法，评价了EQ-5D量表五个维度和两个得分的代答偏倚；（4）基于陕西省第五次国家卫生服务调查数据，比较了3种非倾向评分法以及3种基于复杂抽样的倾向评分法评价中文版EQ-5D的代答效应的效果，重点分析了在PS计算阶段和效应估计阶段，抽样权重和群效应的处理策略。结果：（1）提出了提出了适用于陕西省卫生服务调查复杂抽样特征的分层整群加权校正法；（2）油田职工全人群和男性的高血压和糖尿病的患病率比陕西省居民高，但是女性职工恰好相反，为企业制定健康管理优先政策提供了依据；（3）发现中文版EQ-5D-3L在全人群的调查中，存在显著的代答效应，为大样本调查中EQ-5D代答偏倚的校正提供了依据，同时为后期考虑复杂抽样特征的倾向评分法研究提供比较的基础；（4）大样本调查中，只要混杂因素能够均衡，PS估计阶段是否考虑抽样权重不影响效应估计结果，但效应估计阶段需要考虑抽样权重。复杂抽样多因素分析需要考虑群效应，否则将低估抽样误差。成果：项目资助发表学术论文10篇；其中与研究直接相关的SCI论文1篇，中文2篇。