四种因果图模型在观察性研究因果推断中的比较研究  

A Comparative Study on the Four Causal Diagram Models for Causal Infer-ence in Observational Study

作　　者：覃青连 李峤 颜星星 林越东 周红霞 黄高明 谢志春 唐咸艳 Qin Qinglian;Li Qiao;Yan Xingxing(Department of Epidemiology and Biostatistics,School of Public Health,Guangxi Medical University,530021,Nanning)

机构地区：[1]广西医科大学公共卫生学院流行病与生物统计学系,530021

出　　处：《中国卫生统计》2020年第4期496-500,共5页Chinese Journal of Health Statistics

基　　金：国家自然科学基金(81502890,81960615);美国中华医学基金会(CMB-19-307,CMB-14-202);广西高等学校千名中青年骨干教师培育计划;广西医科大学青年科学基金项目(GXMUYSF201604);广西自然科学基金(2013GXNSFBA019125)。

摘　　要：目的比较有向无环图、结构方程模型、贝叶斯网络和TAN贝叶斯网络四种因果图模型在观察性研究因果推断中的原理方法和应用价值,为因果图模型的合理选用提供参考依据。方法以认知障碍为例,基于先验知识构建轻度认知功能障碍的有向无环图。根据有向无环图建立结构方程模型的初始模型,采用极大似然估计进行参数估计和修正指数进行模型优化。运用爬山算法进行贝叶斯网络结构学习、贝叶斯信息准则进行结构优化和贝叶斯估计进行网络参数学习,并进行网络推理。采用专家建模进行TAN贝叶斯网络的构建,似然比进行独立性测试和极大似然估计进行参数学习,并进行诊断推理。结果实例分析显示,有向无环图、结构方程模型和贝叶斯网络均稳定探测到了结局变量的直接原因且各模型探测到的因果路径基本趋同。有向无环图定性推断了变量间因果关系的概念框架;结构方程模型通过标化路径系数定量推断了模型假定的观测变量与结局变量间的因果关系;贝叶斯网络通过条件概率表定量推断了直接原因组合下结局变量的发生概率,正向预测推理了由因到果的路径关系;TAN贝叶斯网络通过变量重要性评分反向诊断推理了由果到因的路径关系。结论有向无环图、结构方程模型、贝叶斯网络和TAN贝叶斯网络因果图模型在观察性研究因果推断中的侧重点和实际意义有所不同,探测到的因果路径亦有所不同,实际应用时应综合四种因果图模型结果进行因果关系的稳健推断。Objective To compare the principles,methods and applications of directed acyclic graph(DAG),structural equation model(SEM),Bayesian network(BN)and tree augmented naive(TAN)Bayesian network for causal inference in observational study.Methods Taking mild cognitive impairment of elderly for example,four causal diagram models,i.e.DAG,SEM,BN and TAN-BN,were fitted to jointly identify the causal paths and quantitatively infer the strength of causal association between exposure variables and mild cognitive impairment of elderly.Results A total of three exposure variables,i.e.cataract,bowel movement and stroke,were consistently detected as the direct causes of mild cognitive impairment by DAG,SEM and BN.Moreover,causal paths detected by the three causal diagram models were quite identical.Specifically,DAG qualitatively inferred the conceptual framework of causality between exposure variables and outcome variable.SEM quantitatively inferred the causality by the indicator of standardized path-coefficient.BN quantitatively inferred the conditional probability of outcome variable under the combinations of direct causes.While for TAN-BN,the reverse inference of effect-to-cause reasoning was conducted via the indicator of variable importance,and the probability of cause variables underlying the occurrence of outcome variable was quantitatively inferred.Conclusion The practical significance in causal inference for DAG,SEM,BN and TAN-BN is different,yet the causal paths vary somewhat with causal diagram models.Thus,combing the findings of the four causal diagram models is strongly recommended to robustly conduct causal inference.

关 键 词：观察性研究 因果推断 有向无环图 结构方程模型 贝叶斯网络 TAN贝叶斯网络 

分 类 号：O157.5[理学—数学]