目的：在观察性研究或非随机化试验中，由于混杂因素***的存在，研究人员从数据中进行因果推断的能力受到阻碍，本研究利用GBM倾向评分加权法对一组观察性医学数据进行了分析，以期指导相关医学人员进行他们自己的因果推断研究。方法：目前，四类主要的倾向评分法：匹配、分层、逆概率加权和混杂变量调整，已经被普遍用于因果推断的研究。倾向评分法理论上是可以消除可观测到的混杂因素的偏倚，使处理变量接近随机分配设计的效果，从而达到估计处理因素对结局因果效应的目的。结果：考虑到逆概率加权法相对于其它方法的优势，本文概括了它用于因果效应估计的适用条件，特别说明了运用一个现代多元非参数统计技术——广义Boosted模型(GBM)倾向评分加权法的关键环节及优劣。结论：当存在大量不同类型的混杂因素且它们与处理因素之间的线性、非线性或交互效应等函数形式无法确定以及其它问题的时候，GBM倾向评分加权法能克服在精确地估计倾向评分过程中所受到的阻碍，并给出相对更加接近于随机化的因果效应。

Objective In observational studies or non- randomized design, the researchers ?? ability to make causal inferences from data was hampered by confounding factors. This study used this method to analyze a group of observational medical data in order to instruct relevant medical personnel to carry out their own causal inference studies. Methods At present, the four main types of propensity scoring methods: matching, stratification, inverse probability weighting and covariate adjustment have been widely used in the study of causal inference. Propensity score method can theoretically eliminate the bias of the observable confounding factors, so that the treatments variables are close to the result of random assignment design, thus, it is estimated that the treatment factor has a causal effect on the outcome. Results Considering the advantages of the inverse probability weighting method over other methods, this paper summarizes the applicable conditions for the estimate of causal effect, particularly illustrates the use of a modern nonparametric statistical technology-- Generalized Boosted Models (GBM) and its advantages and disadvantages. Conclusion When there is a lot of different types of confounding factors, and uncertain functional forms for their associations with treatment selection in linear, non-linear or interaction effect, and other issues, GBM propensity score weighting method can overcome the obstacles in the process of accurately estimating propensity score.

国家自然科学基金委青年科学基金项目（81502898）：大型观察性医学数据的因果图模型研究，负责人：杨伟；重大新药创制专项子课题（2015ZX09501004-001-007）：临床需长期使用的中药口服制剂安全性监测科研研究，负责人：李学林。