基于g-computation联合混合效应模型控制未测混杂因素的因果推断方法模拟研究及实例验证  

Simulation Study and Case Validation on Causal Inference of g-computation-based Joint Mixed-effects Model for Controlling Unmeasured Confounders

作　　者：孙博然 芦文丽[1] 陈永杰 Sun Boran;Lu Wenli;Chen Yongjie(Department of Epidemiology and Health Statistics,School of Public Health,Tianjin Medical University,Tianjin 300070)

机构地区：[1]天津医科大学流行病与卫生统计学系,300070

出　　处：《中国卫生统计》2024年第5期691-696,共6页Chinese Journal of Health Statistics

基　　金：国家自然科学基金青年项目(81903416)。

摘　　要：目的通过模拟实验和实例验证探讨基于g-computation的联合混合效应模型(joint mixed-effects model,JMM)控制纵向研究未测混杂因素进行因果推断时的效果及性能特点。方法通过计算机模拟产生包含基线及两次随访时点的纵向数据,模拟条件包括样本含量、有无未测混杂因素及未测混杂效应大小,分别利用基于g-computation的JMM、线性混合效应模型、固定效应模型和纵向目标极大似然估计方法估计因果效应,通过平均绝对偏差(mean absolute deviation,MAD)、标准误、均方根误差(root mean square error,RMSE)、95%置信区间覆盖率(95%confidence interval coverage,95%CI coverage)评价比较各方法因果推断的效果。利用绝经期女性队列体检数据,应用四类模型分别估计绝经期女性血清卵泡刺激素(follicle-stimulating hormone,FSH)水平与腰椎骨密度间因果关系,对各模型在真实纵向数据中的因果推断效果进行验证。结果JMM控制未测混杂因素的因果推断准确性最佳,但稳定性略差。当研究中存在较强未测混杂效应时,仅JMM可准确估计因果效应,且其在大样本量时估计的精确性和真实性较好。结论基于g-computation的JMM可有效控制纵向研究中未测混杂因素进行近似无偏因果推断。Objective A simulation study was conducted to explore the effect and performance of g-computation-based joint mixed-effects model(JMM)on causal inference for controlling unmeasured confounders in longitudinal studies.Methods Longitudinal data including baseline and two follow-up visits were generated by computer simulations.The simulation scenarios included different sample sizes,the presence or absence of unmeasured confounders,and effects of unmeasured confounders.Causal effects were estimated using g-computation-based JMM,linear mixed-effects model,fixed effects model,and longitudinal target maximum likelihood estimation,respectively.Indicators including mean absolute deviation(MAD),standard error,root mean square error(RMSE),and 95%confidence interval coverage(95%CI coverage)were used to evaluate and compare the causal inference performance.Based on the physical examination cohort data of the menopausal women,four models were used to estimate the causal association between serum follicle-stimulating hormone(FSH)levels and lumbar bone density in menopausal women respectively,verifying the causal inference performance of models in the real longitudinal data.Results JMM had a better accuracy of causal inference with controlling unmeasured confounders.But its estimation stability was slightly worse.When strong unmeasured confounders existed,only JMM can accurately estimate the causal effect,and its precision and authenticity were better in scenarios with large sample sizes.Conclusion JMM can effectively control the unmeasured confounders and perform approximately unbiased causal estimation in longitudinal studies.

关 键 词：纵向研究 未测混杂因素 g-computation 联合混合效应模型 

分 类 号：R195.1[医药卫生—卫生统计学]