Understanding causal mechanisms in complex systems requires evaluating path-specific effects (PSEs) in multi-mediator models. Identification of PSEs traditionally relies on the demanding cross-world independence assumption. To relax this, VanderWeele et al. (2014) proposed an interventional approach that redefines PSEs, while Stensrud et al. (2021) introduced dismissible component conditions for identifying separable effects under competing risks. In this study, we leverage SWIGs to dissect the causal foundations of these three semantics and make two key advances. First, we generalize separable effects beyond competing risks to the setting of multi-mediator models and derive the assumptions required for their identification. Second, we unify the three approaches by clarifying how they interpret counterfactual outcomes differently: as mediator-driven effects (classical), randomized contrasts (interventional), or component-specific actions (separable). We further demonstrate that violations of cross-world independence originate from mediators omitted from the model. By analogy to confounder control, we argue that just as exchangeability is achieved by conditioning on sufficient confounders, cross-world independence can be approximated by including sufficient mediators. This reframing turns an abstract assumption into a tangible modeling strategy, offering a more practical path forward for applied mediation analysis in complex causal systems.

翻译：理解复杂系统中的因果机制需要评估多中介模型中的路径特定效应。传统上，PSE的识别依赖于苛刻的跨世界独立性假设。为放宽此假设，VanderWeele等人（2014）提出了一种重新定义PSE的干预方法，而Stensrud等人（2021）则在竞争风险场景下提出了用于识别可分离效应的可忽略成分条件。本研究利用SWIGs剖析了这三种语义的因果基础，并取得两项关键进展。首先，我们将可分离效应从竞争风险推广至多中介模型，并推导出其识别所需的假设条件。其次，我们通过阐明三种方法对反事实结果的不同解释——中介驱动效应（经典方法）、随机化对比（干预方法）或成分特异性干预（可分离方法），实现了三者的理论统一。我们进一步证明，跨世界独立性的违背源于模型中遗漏的中介变量。类比于混杂因素控制，我们论证道：正如通过调整充分混杂因素可实现可交换性，纳入充分的中介变量亦可近似满足跨世界独立性。这一重述将抽象假设转化为具体的建模策略，为复杂因果系统中的应用中介分析提供了更具实践意义的研究路径。