目标最大似然估计在药品安全性主动监测中的应用  

Application of Targeted Maximum Likelihood Estimation in Active Drug Safety Surveillance

作　　者：韩贺东[1] 郭威[1] 叶小飞[1] 许金芳[1] 郭晓晶[1] 朱田田[1] 史文涛[1] 王蒙[1] 侯永芳[2] 贺佳[1] 

机构地区：[1]第二军医大学卫勤系卫生统计学教研室,上海200433 [2]国家食品药品监督管理总局药品评价中心,北京100045

出　　处：《中国药物警戒》2017年第6期341-345,共5页Chinese Journal of Pharmacovigilance

基　　金：国家自然科学基金(81373105):贝叶斯倾向性评分方法研究及其在药品不良反应信号检测中的应用;上海市公共卫生重点学科(815GWZK0901):循证公共卫生与卫生经济学;上海市卫计委科研课题(201440379):药品不良反应信号检测中混杂因素控制的倾向性评分方法研究

摘　　要：目的利用观察性数据进行因果效应的推断一直是研究者关注的焦点。在药物流行病学领域,探索药物疗效或不良反应也需要因果推断。由van deer laan提出的目标最大似然估计(targeted maximum likelihood estimation,TMLE)被证明具有良好的特性,本文对该方法的原理及应用现状做一介绍。方法因果推断常用的方法包括逆概率加权法,G-算法(G-formulation)及一些双稳健的估计方法。通过文献检索,对TMLE法的原理、模型构建、效应估计、统计推断、性质及应用等方面进行回顾与综述,同时比较TMLE法与其他几种因果推断方法的异同。结果与其他因果推断方法相比,TMLE法具有一定的优势,然而其实现过程更复杂。结论 TMLE法是一种双稳健的因果推断方法,能产生目标参数的有效无偏估计,可为TMLE法在药品不良反应主动监测因果推断中的应用提供参考,以期进一步加强药品风险管理。Objective Causal inference using observational data has long been the focus of epidemiologic researchers. In pharmacoepidemiologic studies, exploring effect of a drug or adverse drug reaction caused by a drug also fits into causal inference. Targeted maximum likelihood estimation （TMLE） proposed by van deer laan is proved of desirable properties, we introduce its principle and application in the study. Method The common methods for causal inference include inverse probability of treatment weighting （IPTW）, G-formulation and some double robust estimators. We provide an overview of the theory, model, estimation, statistical inference and properties. Meanwhile, we also compare the performances of several causal effects estimators. Result Compared with other methods, TMLE has certain advantages. However, its implementation is difficult. Conclusion TMLE is a method with double robustness and could also produce efficient unbiased estimator of targeted parameter. There is potential value of TMLE enhancing drug risk management in active drug safety surveillance.

关 键 词：目标最大似然估计 因果推断 药物流行病学 药品不良反应主动监测 

分 类 号：R994.11[医药卫生—毒理学]