二分类数据缺失多重填补分析及应用  

The Multiple Imputation and Application in Binary Longitudinal Missing Data

作　　者：张耀[1] 陈培翠 张翠仙[1] 罗天娥[1] 刘桂芬[1] 

机构地区：[1]山西医科大学公共卫生学院卫生统计教研室,030001

出　　处：《中国卫生统计》2014年第3期370-373,共4页Chinese Journal of Health Statistics

基　　金：国家自然科学基金项目(编号81172774);国家青年科学基金项目资助(81001294);太原市大学生创新创业专题(120164023)

摘　　要：目的阐明四种填补方法(multiple imputation,MI)的基本原理,实例介绍纵向研究二分类缺失数据多种填补方法的应用。方法对比分析简单填补、分层填补、考虑个体差异的填补及考虑个体、抽样的多重填补等四种填补方法;模拟证实几种OR取值的敏感性分析。结果进行大样本(N=10000)模拟研究表明:简单多重填补分析会降低检验效能,不能客观反应两样本的差异;考虑先前信息的分层多重填补会扩大I型错误;若只考虑个体变异,仅模拟一个数据集,所得结论不稳定;在考虑个体、抽样和填补差异后模拟的多重填补数据集,当OR≈2时,所得统计量基本接近真值;实例验证,经高血压知晓干预后,尚不能认为两区的吸烟率有差别。结论不考虑前次观察数据以及OR值的影响,一味地把缺失值当作该事件发生处理,会加大I型错误;只有综合考虑个体、抽样和填补差异,多重填补数据集的估计结果才更具稳健性。Objective To clarify the basic principles of the multiple imputation （ MI）, we will introduce several methods with examples. Methods Compare the analysis of four MI model, i. e. （ 1 ） simple MI. （2） Stratified MI. （ 3 ） The MI which consider individual differences. （ 4 ） Perform the comprehensive analysis considering the individual, sampling and imputation. Carry out sensitivity analysis under different imputation sample,using SAS 9. 2 to complete MI. Results Large sample（N = 10000） simulation show that. simple multiple imputation analysis will reduce the ability of performance test, it can not response the difference between two samples, the multiple imputation analysis which considering the previous information will expand type I error. If only considerate the individual variability and simulate a data set, the conclude will be not stable; considerate the individual variability, sampling, and filling difference, when OR ≈ 2, the statistics result are close to the true value. We finally still can not believe that the rate of smoking are unequal between the two areas though the example of hypertension awareness intervention. Conclusion When we regard the missing as the event, there will increase the probability of type I error. When we consider the difference of individual, sampling and multiple imputation, we will draw a more robust parameter estimation.

关 键 词：多重填补 纵向研究 二分类数据缺失 效果评价 

分 类 号：R195[医药卫生—卫生统计学]