有向无环图在混杂因素识别与控制中的应用及实例分析

观察性研究是流行病学病因研究常用的研究设计,但应用观察性研究进行因果推断时,常由于未经识别、校正的混杂因素的存在,歪曲暴露因素与研究结局之间的真实因果关系。传统混杂因素判断标准在实际应用中不够直观,且有一定局限性,有时甚至出现混杂因素的误判。有向无环图(DAGs)可以直观识别观察性研究中存在的混杂因素,将复杂的因果关系可视化,判断研究中需要校正的最小校正子集,并可避免传统混杂因素判断标准的局限性,结合DAGs还可以指导混杂因素校正方法的选择,在观察性研究中因果推断具有重要指导价值,DAGs在未来的流行病学研究中将有更多的应用。     

在因果推断中应用有向无环图识别和控制选择偏倚

在流行病学研究中，选择偏倚会导致研究样本无法代表一般人群，使研究结果偏离真实值，无法推断真实的因果关联。本文通过构建有向无环图（directed acyclic graphs，DAGs），将复杂的因果关系可视化，提供识别选择偏倚的直观方法，并通过冲撞分层偏倚的图形结构来验证不同类型的选择偏倚。在实际研究中，可能同时存在多种偏倚，对冲撞变量进行不恰当的调整会新增冲撞分层偏倚，打开后门路径，引入混杂偏倚，甚至改变原有混杂偏倚的大小与方向。为了得到暴露到结局的无偏估计，研究者可以通过构建DAGs，帮助识别冲撞变量，防止冲撞偏倚的发生。     

客观世界的因果关系:基于有向无环图的结构解析

客观世界因果关系的整体框架较为笼统而缺乏明晰的细节,给因果关系的研究带来困难。本文基于因果关系的时序特性结合有向无环图（DAGs）,以因和果的发生时间为界,将客观世界的时间维度划分为3个时间域和2个时间点。通过对5个时间单位上变量间存在着的完整的因果关系的病因网络DAGs进行分析发现,其病因结构由两部分叠加组成：第一部分是各个时间域间/时间点上任取一变量间的组合DAGs,为因果关系的基本结构,构成病因网络的核心,仅混杂路径影响其因果效应估计;第二部分是各个时间域内/时间点上变量间的母子或祖先-后代关系,其DAGs表现为与混杂类似的结构。本文简洁明了地构建了客观世界因果关系研究的整体框架（病因网络DAGs）,解释了控制混杂以解决因果效应估计的结构基础,为正确研究和识别因果关系奠定基础。     

匹配在观察性研究中的作用——有向无环图视角

匹配是观察性研究中选择研究对象的一种常用方法,具有控制混杂因素、提高统计效率等作用,但其控制混杂因素的作用在不同观察性研究中并不一致,匹配在队列研究中能够消除匹配变量的混杂偏倚,但在病例对照研究中匹配本身并不能消除混杂偏倚。在匹配性病例对照研究选择匹配变量时,研究者可能并不能准确判断该变量是否为混杂变量,若误将真实情况为非混杂因素的变量进行匹配,则会形成过度匹配,造成统计效率下降或引入难以避免的偏倚或增加工作量等;若将真实情况为混杂因素的变量遗漏不予匹配,则会造成混杂偏倚。有向无环图是一种直观的展示不同流行病学研究设计、变量间复杂因果关系的可视化图形语言。本文从有向无环图视角分析匹配在不同观察性研究设计中的作用、匹配性病例对照研究中欲匹配变量的选择标准制定,为今后流行病学研究设计提供一定的参考建议。     

四种因果图模型在观察性研究因果推断中的比较研究

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。     

研究设计时混杂控制策略的结构分类

混杂影响着人群因果关系的发生。依据混杂因素是否已知、可测量及已测量,可将其分为4类情形。基于有向无环图,对混杂的控制策略分为两类：①混杂路径打断法,又可分为单路径和双路径打断法,分别对应于暴露完全干预法、限制法和分层法;②混杂路径保留法,分别对应于暴露不完全干预法（工具变量设计或不完美的随机对照试验）、中间变量法和匹配法。其中,随机对照试验、工具变量设计或孟德尔随机化设计、中间变量分析可满足4类混杂的控制,而限制法、分层法和匹配法仅适用于已知、可测量并已测量的混杂。识别不同类型混杂的控制机制,有助于在研究设计阶段提出应对措施,是获得正确因果效应估计的前提。     

因果图模型及其在营养流行病学研究中的应用

不良饮食是慢性非传染性疾病最重要的可控危险因素之一,但通过随机对照试验定量阐明具体饮食因素与健康结局的因果关联面临很多困难。近年来,因果推断的迅速发展为充分利用和发掘观察性研究数据,产生高质量的营养流行病学研究证据提供了有力的理论和方法工具。其中,因果图模型通过整合大量先验知识将复杂的因果关系系统可视化,提供了识别混杂和确定因果效应估计策略的基础框架,基于不同的因果图,可选择调整混杂、工具变量或中介分析等不同的分析策略。本文对因果图模型的思想和各种分析策略的特点及其在营养流行病学研究中的应用进行介绍,旨在促进因果图模型在营养领域的应用,为后续研究提供参考和建议。     

“探针变量”：观察性研究中混杂因素识别的工具

观察性研究中往往存在未知或未测量的混杂因素，是流行病学因果关联研究中的重大挑战。本文介绍一种可以应用在观察性研究中的一种对未知/未测量混杂因素进行识别和效应评估的工具——“探针变量”。其主要可以分为暴露探针变量、结局探针变量以及中介探针3种形式，前2种不仅可以对未知/未测量混杂因素进行识别，也可以对其效应量进行估计，从而揭示真实的暴露与结局之间的关联。而中介探针则是针对“中介因子”进行控制，从而识别暴露和结局之间是否存在未测量混杂因素。该理论实践过程中最大的困难在于“探针变量”的选择和确定，不恰当的“探针变量”可能引入新的混杂，导致未测量混杂因素识别不准确。“探针变量”可以推荐作为观察性研究报告中的一项敏感性分析内容，有助于读者真实理解暴露与结局之间的关联，增加观察性流行病学研究中的证据力度。     

因果中介效应分析及SAS实现

目的 通过模拟数据，研究父母提供的鼓励环境是否会对儿童的认知发展产生影响，探索父母鼓励是否增强了孩子的学习动机，介绍因果中介效应分析方法的原理及SAS实现。方法 在未控制混杂因素和控制混杂因素两种情况下，运用因果中介效应分析方法对鼓励和认知得分之间的因果路径进行分解，确定中介变量动机在因果关系中的作用程度。结果 学习动机在父母鼓励与儿童认知发展之间起中介作用，中介效应占总效应的比例为47%(不控制混杂因素)、37%(控制混杂因素)。结论 学习动机是中介变量，父母鼓励可以通过增强孩子的学习动机来提高儿童的认知发展。在满足相关前提和假设下，CAUSALMED过程可以实现因果中介效应分析，探索因果关系的内在影响机制。     

基于有向无环图和线性回归理论定量分析常见混杂的影响

目的结合有向无环图(directed acyclic graphs, DAGs)与线性回归模型,提出常见混杂的定量分析方法。方法针对典型的两种DAGs(情形1:X←C→Y;情形2:X←C_1→M←C_2→Y),基于线性回归理论给出偏倚大小的定量表达式,并探讨各参数对因果效应估计的影响。结果对于情形1的DAG,暴露X与结局Y的因果效应估计需通过关闭二者间后门路来控制混杂C,否则X与Y间因果估计会受到混杂偏倚的影响。理论推导结果显示,若只改变C对X的效应则同时会增加X的方差,此效应的强弱对混杂偏倚的影响是非单调的,除非在改变C与X的效应的同时控制X的方差不变,而C与Y间的效应强弱对混杂偏倚的影响则是单调的。对于情形2,又称M-DAGs无需做变量控制即可做X与Y的因果估计,但当错误地控制碰撞点M后会导致后门路径打开而出现混杂,此时需进一步控制C_1和(或)C_2来关闭后门路径。我们用回归理论解释了该结果,并且得到当C_1与X相关性较高时,同时控制C_1和M的方法结果会不稳定。结论本研究基于有向无环图,根据线性回归理论推导出常见混杂的定量分析方法,该方法也适用于无法观测的混杂,为多因素因果推断提供了一种实用工具。     

Causal directed acyclic graphs and the direction of unmeasured confounding bias

We present results that allow the researcher in certain cases to determine the direction of the bias that arises when control for confounding is inadequate. The results are given within the context of the directed acyclic graph causal framework and are stated in terms of signed edges. Rigorous definitions for signed edges are provided. We describe cases in which intuition concerning signed edges fails and we characterize the directed acyclic graphs that researchers can use to draw conclusions about the sign of the bias of unmeasured confounding. If there is only one unmeasured confounding variable on the graph, then nonincreasing or nondecreasing average causal effects suffice to draw conclusions about the direction of the bias. When there are more than one unmeasured confounding variable, nonincreasing and nondecreasing average causal effects can be used to draw conclusions only if the various unmeasured confounding variables are independent of one another conditional on the measured covariates. When this conditional independence property does not hold, stronger notions of monotonicity are needed to draw conclusions about the direction of the bias.     

Estimating and testing direct genetic effects in directed acyclic graphs using estimating equations

In genetic association studies, it is important to distinguish direct and indirect genetic effects in order to build truly functional models. For this purpose, we consider a directed acyclic graph setting with genetic variants, primary and intermediate phenotypes, and confounding factors. In order to make valid statistical inference on direct genetic effects on the primary phenotype, it is necessary to consider all potential effects in the graph, and we propose to use the estimating equations method with robust Huber–White sandwich standard errors. We evaluate the proposed causal inference based on estimating equations (CIEE) method and compare it with traditional multiple regression methods, the structural equation modeling method, and sequential G‐estimation methods through a simulation study for the analysis of (completely observed) quantitative traits and time‐to‐event traits subject to censoring as primary phenotypes. The results show that CIEE provides valid estimators and inference by successfully removing the effect of intermediate phenotypes from the primary phenotype and is robust against measured and unmeasured confounding of the indirect effect through observed factors. All other methods except the sequential G‐estimation method for quantitative traits fail in some scenarios where their test statistics yield inflated type I errors. In the analysis of the Genetic Analysis Workshop 19 dataset, we estimate and test genetic effects on blood pressure accounting for intermediate gene expression phenotypes. The results show that CIEE can identify genetic variants that would be missed by traditional regression analyses. CIEE is computationally fast, widely applicable to different fields, and available as an R package.     

Competing risk bias to explain the inverse relationship between smoking and malignant melanoma

The relationship between smoking and melanoma remains unclear. Among the different results is the paradoxical finding that smoking was shown to be inversely associated with the risk of malignant melanoma in some large cohort and case-control studies, even after control for suspected confounding variables. Smoking is a known risk factor for many non-communicable diseases, including coronary heart disease, stroke, as well as other malignancies; it has been shown to be positively associated with other types of skin cancer, and there remains no clear biologic explanation for a possible protective effect on malignant melanoma. In this paper, we propose a plausible mechanism of bias from smoking-related competing risks that may explain or contribute to the inverse association between smoking and melanoma as spurious. Using directed acyclic graphs for formalization and visualization of assumptions, and Monte Carlo simulation techniques, we demonstrate how published inverse associations might be compatible with selection bias resulting from uncontrolled or unmeasured common causes of competing outcomes of smoking-related diseases and malignant melanoma. We present results from various scenarios assuming a true null as well as a true positive association between smoking and malignant melanoma. Under a true null assumption, we find inverse associations due to the biasing mechanism to be compatible with published results in the literature, especially after the addition of unmeasured confounding variables. This study could be seen as offering a cautionary note in the interpretation of published smoking-melanoma findings.     

有向无环图:语言、规则及应用

几乎所有的科学研究都在探索因果关系,有向无环图（DAGs）是因果关系研究的图形工具。本文系统地介绍了DAGs的图形语言、基本规则和干扰规则,及其在识别研究问题、理解和实施研究设计、指导数据分析、偏倚分类等方面的应用。DAGs对因果关系的研究具有重要的指导价值。     

Directed acyclic graphs helped to identify confounding in the association of disability and electrocardiographic findings: results from the KORA-Age study

ObjectivesTo examine the association between electrocardiographic (ECG) findings and disability status in older adults.Study Design and SettingKORA-Age, a population-based cross-sectional study of the MONICA/KORA project, a randomized sample from Southern Germany of people aged 65 years or older.ResultsA total of 534 (51.5%) of 1,037 participants were characterized as disabled. Disabled participants were on average 4.5 years older than those who were not disabled. Crude associations of left-axis deviation, ventricular conduction defects, atrial fibrillation, and QT prolongation with disability status were significant (P < 0.05). In models controlled for age and sex, these effects remained constant except for QT prolongation. In the models adjusted for the minimal sufficient adjustment set (consisting of the variables sex, physical activity, age, obesity, diabetes, education, heart diseases, income, lung diseases, and stroke) identified by a directed acyclic graph (DAG), no significant association could be shown.ConclusionAssociations between specific ECG findings and disability were found in unadjusted analysis and logistic models adjusted for age and sex. However, when adjusting for other possible confounders identified by the DAG, all these associations were no longer significant. It is important to adequately identify confounding in such settings.     

Four types of effect modification: a classification based on directed acyclic graphs

It is possible to classify the types of causal relationships that can give rise to effect modification on the risk difference scale by expressing the conditional causal risk-difference as a sum of products of stratum-specific risk differences and conditional probabilities. Directed acyclic graphs clarify the causal relationships necessary for a particular variable to serve as an effect modifier for the causal risk difference involving 2 other variables. The directed acyclic graph causal framework thereby gives rise to a 4-fold classification for effect modification: direct effect modification, indirect effect modification, effect modification by proxy and effect modification by a common cause. We briefly discuss the case of multiple effect modification relationships and multiple effect modifiers as well as measures of effect other than that of the causal risk difference.     

Quantifying biases in causal models: classical confounding vs collider-stratification bias

It has long been known that stratifying on variables affected by the study exposure can create selection bias. More recently it has been shown that stratifying on a variable that precedes exposure and disease can induce confounding, even if there is no confounding in the unstratified (crude) estimate. This paper examines the relative magnitudes of these biases under some simple causal models in which the stratification variable is graphically depicted as a collider (a variable directly affected by two or more other variables in the graph). The results suggest that bias from stratifying on variables affected by exposure and disease may often be comparable in size with bias from classical confounding (bias from failing to stratify on a common cause of exposure and disease), whereas other biases from collider stratification may tend to be much smaller.