GEE、GLMM和MLM分析卫生重复测量资料的效果比较  

Effect comparison of GEE,GLMM,and MLM in analyzing repeated measures of health-related field

作　　者：周婷[1] 兰蓝[1] 邱建青 杜春霖[1] 李晓松[1] 张韬[1] 

机构地区：[1]四川大学华西公共卫生学院(华西第四医院),四川成都610041

出　　处：《现代预防医学》2017年第16期2881-2885,2899,共6页Modern Preventive Medicine

基　　金：国家自然科学基金青年科学基金项目(No.81602935);四川大学青年教师科研启动基金(2016SCU11006)

摘　　要：目的探讨广义估计方程(Generalized Estimating Equation,GEE)、广义线性混合模型(Generalized Linear Mixed Models,GLMM)和多水平统计模型(Multilevel statistical Models,MLM)分析重复测量资料的效果。方法采用SAS9.4和MLwi N2.26对住院病人人均医药费影响因素拟合GEE、GLMM和MLM,其中GEE设置独立型(type=ind)、可交换型(type=exch)、自相关型(type=ar)和非确定结构型(type=un)4种作业相关矩阵,比较效应估计值标准误的差异。结果 (1)纳入研究的变量有随时间增加的趋势;(2)GEE(type=ind)和GLMM模型下各变量效应估计值标准误具有一致性,时间依存变量的效应估计值标准误被高估,时间独立变量的标准误被低估;(3)MLM和GEE(type=exch)模型下各变量效应估计值标准误具有一致性,对研究的重复测量资料有较好的反映;(4)GEE(type=ar)估计效果居于以上二者之间,GEE(type=un)未收敛。结论卫生相关资料具有重复测量资料特征,统计分析需要考虑资料内部相关结构,对此类资料MLM与GEE(type=ind)、GLMM与GEE(type=ind)分别有相同效果。Objective The study aimed to explore the effect of repeated measurement data of generalized estimating equation （GEE）, generalized linear mixed models （GLMM）, and multilevel statistical models （MLM）. Methods GEE, GLMM, and MLM were conducted for influencing factors of medical expenses for inpatients by adopting SAS9.4 and MLwiN2.26. For GEE, we set four types of the working independence correlation model, independence （type =ind）, exchangeable （type =exch）, autoregressive （type=at）, and unstructured （type=un）. The standard error （SE） was compared. Results 1. Each variable included in the model has a tendency to increase over time. 2. For GEE（type=ind） and GLMM, SE of all the variables showed good consistence and SE of time-dependent explanatory variables was overestimated, while time-independent explanatory variables were underestimated. 3. For MLM and GEE （type=exch）, SE of all the variables showed good consistence and the estimated effects were approached to true situation. 4. The estimated effect of GEE （type =at） was in between the two categories and the results of GEE （un） was without the model converging. Conclusion Health-related data has the characteristics of repeated measurement data, of which intra-class correlation structure should be considered in statistical analysis. And the two categories of MLM and GEE （type=exch）, GLMM and GEE （type=ind） had similar effects, respectively.

关 键 词：重复测量资料 广义估计方程 广义线性混合模型 多水平统计模型 

分 类 号：R181.2[医药卫生—流行病学]