几种非线性混杂变量校正方法的性能比较  

Performance Comparisons of Several Adjustment Methods for Nonlinear Confounding Variable

作　　者：巩浩雯 熊殷 刘玉秀 陈文松 许敏怡 刘曼 李维勤[5] Gong Haowen;Xiong Yin;Liu Yuxiu(Department of Biostatistics,School of Public Health,Southern Medical University(510515),Guangzhou)

机构地区：[1]南方医科大学公共卫生学院生物统计学系,510515 [2]东部战区总医院重症医学科数据与统计室 [3]南京医科大学金陵临床医学院 [4]南京医科大学公共卫生学院生物统计学系 [5]南京大学医疗健康大数据国家研究院

出　　处：《中国卫生统计》2023年第3期326-330,共5页Chinese Journal of Health Statistics

基　　金：国家自然科学基金面上项目(81473066)。

摘　　要：目的比较临床观察性研究中几种非线性连续性混杂变量校正方法的统计性能,为准确评价暴露因素影响的校正方法选择提供依据。方法基于两分类结局变量logistic回归模型,设定不同复杂程度的自变量关系,进行两分类暴露因素的效应量OR值估计。通过Monte-Carlo模拟研究,采用OR的相对偏倚、95%置信区间覆盖率,对比评价将非线性连续混杂变量使用线性校正、分类校正、多项式校正、分数多项式校正、以及使用限制性立方样条、广义加性模型校正后对暴露效应估计的影响。使用Kaggle数据平台中抑郁症数据,对运动与抑郁症患病风险之间的因果关系进行了实例分析。结果连续混杂变量为非线性时,在几种校正方法中,仅使用线性校正会造成效应估计的较大偏倚,使用限制性立方样条校正和广义加性模型(光滑函数为薄板样条)得到的估计值偏倚最小。实例分析结果显示,不考虑非线性混杂校正将导致效应量估计的偏倚。结论评价暴露因素的影响时,需要考虑非线性混杂变量的校正问题,推荐选用限制性立方样条或广义加性模型进行校正分析。Objective To compare the statistical performance of several correction methods for nonlinear continuous confounding variables in clinical observational studies and to provide a basis for the selection of methods to accurately evaluate the effects of exposure factors.Methods Based on logistic regression model with different levels of complexity of the variable relationships,the estimation of OR values for the effect sizes of two-category exposure factors was performed.We carried out a simulation study to investigate the effect of adjusting for a nonlinear confounder in the estimation of relationship between the exposure and outcome in 7 ways:linear,categorical,polynomial,fractional polynomial,restricted cubic splines and generalized additive model.An example analysis of the causal relationship between exercise and risk of depression was conducted using depression data from Kaggle.Results When nonlinear continuous confounding variables exist,linear correction results in a large bias,while restricted cubic spline correction and generalized additive model(smooth function as a thin slab spline)yield a less biased estimate.The results of the empirical analysis showed that not considering the nonlinear confounding correction would lead to biased estimates.Conclusion When evaluating the effects of exposure factors,correction for nonlinear confounding variables needs to be considered,and the restricted cubic spline or generalized additive models are recommended for analysis.

关 键 词：非线性变量 混杂 限制性立方样条 广义加性模型 MONTE-CARLO模拟 

分 类 号：R195.1[医药卫生—卫生统计学]