复杂抽样数据多水平模型分析方法及其应用  

Application of Multilevel Modeling to Complex Sample Survey Data

作　　者：于石成[1] 廖加强[2] 于妺 郭莹[1] 肖革新[1] 金承刚[3] 冯国双[1] 胡跃华[1] 马林茂[1] 

机构地区：[1]中国疾病预防控制中心公共卫生监测与信息服务中心,102206 [2]四川大学华西公共卫生学院卫生统计教研室,610041 [3]北京师范大学社会发展与公共政策学院,100875

出　　处：《中国卫生统计》2014年第2期193-196,201,共5页Chinese Journal of Health Statistics

基　　金：淮河流域癌症综合防治项目(1310800003)

摘　　要：目的本文通过抽样调查实例,阐述多阶段抽样、不等抽样概率和事后分层特性不同产生的复杂抽样数据,其应用多水平模型分析的原理和方法。方法对我国某省行为危险因素抽样调查的数据,应用未加权和加权的随机截距logistic回归模型分析了某些因素与跌倒性伤害的关系。结果实际分析包括50个区县(PSU),250个乡镇街道(2水平),12086个体(1水平)。未加权估计结果显示:对跌倒性伤害有统计学影响的变量是健康状况中等和差、未被雇佣和未婚,年龄为负相关,即年龄越大,发生跌倒性伤害的危险性越小;复杂抽样2水平logistic回归分析显示:对跌倒性伤害有统计学影响的变量与未加权的结果基本一致,但未婚失去了统计学意义。体重指数、性别和受教育程度与跌倒性伤害的发生没有统计学联系。结论与未加权的结果比,加权分析对跌倒性伤害有统计学影响的变量基本一致,但加权复杂抽样PMLE估计的标准误偏大,结果更保守;对性别的分析发现,加权后的结果符合目前对跌倒性伤害发生机制的认识,因此纳入权重的多水平分析方法对该资料可能更合理。Objective To illustrate the principal and application of multilevel modeling of complex survey data that were derived from multistage sampling, unequal sampling probabilities and different features of post-stratification. Methods Weighted and un-weighted random intercept logistic regression models were applied to complex survey data of behavioral risk factors in a province to look at the asso- ciation of fall injuries with some factors of interest. Results There were 12086 subjects （level 1 ） aged 45 years or above nested within 250 villages, towns and sub-districts （ level 2） from 50 counties/districts （PSU）. Un-weighted results showed that variables significantly and positively associated with the risk of fall injuries were fair or poor health, unemployed situation, unmarried; age was significantly and negatively associated with the risk of fall injuries, or one less likely got injured when getting older. The results from 2-level random inter- cept logistic model demonstrated that the variables associated with the risk of fall injuries were similar to those from un-weighted models, but the variable of unmarried mitigated its significance to be insignificant. Body mass index, being male, educational level were not associ- ated with the risk of fall injuries from the analyses. Conclusion In contrast to the results from un-weighted methods, statistically signifi- cant variables from weighted methods were analogous to those from weighted ones; however, estimates using full pseudo-maximum-likeli- hood estimation （PMLE） were more conservative as opposed to un-weighted ones. As for gender, weighted result was in consistent with the current understanding of the mechanism for the development of fall injuries, therefore, it sounded more reasonable to employ multilevel modeling for the complex survey data.

关 键 词：复杂抽样 多水平模型 多阶段抽样 随机效应logistic回归 

分 类 号：R195[医药卫生—卫生统计学]