回归校准法对线性回归中信息偏倚校正的模拟研究  

A Simulation Study on Regression Calibration to Correct Measurement Error in Linear Regression Model

作　　者：张桥[1] 张秋菊[1] 陈霞[1] 李慧婷[1] 陈永杰[1] 杨建飞[1] 刘美娜[1] 

机构地区：[1]哈尔滨医科大学公共卫生学院,150086

出　　处：《中国卫生统计》2013年第3期327-329,共3页Chinese Journal of Health Statistics

基　　金：国家科技支撑计划(2011BAI09B02)

摘　　要：目的探讨回归校准法对简单线性回归中,符合正态分布或均匀分布的信息偏倚校正效果。方法通过SAS软件模拟产生有信息偏倚的数据集,用回归校准法对信息偏倚进行校正,然后对比设定的真实β值和校正后的β*值之间的差别,来评价校正效果。结果无论信息偏倚符合正态分布还是均匀分布,回归校准法都能达到较好的校正效果;在固定信息偏倚大小和验证数据集样本量,但不同β值的情况下,信息偏倚校正效果均较好;未校正时的β值都小于设定的真实β值,且信息偏倚越大,β值越小;随着验证数据集样本量的增加,回归校准法对信息偏倚的校正效果越好。结论可以使用回归校准法对线性回归中符合正态分布或均匀分布的信息偏倚进行校正;在校正时要根据信息偏倚的大小来确定合适的验证数据集样本量,以最少验证数据集样本量达到最佳校正效果;在有信息偏倚存在的线性回归分析中,一定要对回归系数进行校正。Objective To study the correction effect of regres- sion calibration when correcting measurement error, which meets normal distribution or uniform distribution, in linear regression model. Methods At first, we simulated datasets that contain measurement error by SAS soft- ware package, and used regression calibration to correct measurement error. Then we evaluated the correction effect by comparing the difference be- tween true value and corrected value. Results Regression calibration method can achieve good correction effect, no matter the measurement error meets normal distribution or uniform distribution. When fixed the measure-ment error and the validation data set sample size, the correction effects were all perfect with different β values. All the uncorrected values lower than the true values,the greater measurement error,the smaller value. Along with the increase of the validate data set sample size, the correction effect appears better. Conclusion Using regression calibration to correct meas- urement error, which meets normal distribution or uniform distribution, in linear regression model can have a good correction effect. In order to a- chieve the best correction effect with the minimum validate dataset sample size,we should determine an appropriate validate dataset sample size ac- cording to the size of the measurement error. If a linear regression model has measurement error, it is necessary to be corrected by using the regres- sion calibration.

关 键 词：信息偏倚 回归校准法 线性回归 回归稀释 

分 类 号：R195[医药卫生—卫生统计学]