Mechanisms, Capacities, and Dependencies: A New Theory of Causal Reasoning

因果认知赋予人类令人印象深刻的能力。它使我们能够预测和解释，计划行动，理解混乱，对转移情况做出有根据的猜测，思考反事实的替代方案，以及使用和设计创新设备。目前还没有心理学理论能够在一个完整的框架内提供所有这些能力的完整说明。此外，侧重于具体任务的部分理论往往忽视能力之间的相互作用。另一个缺点是大多数理论忽略或低估了机制背景知识的重要作用。在过去的几年里，哲学和计算机科学领域提出了一些更先进的理论，承认了其中的一些缺陷。最近的一个发展是珀尔的因果关系理论。珀尔改变了他先前认为因果关系可以归结为协变的观点。他现在认为，抽象的结构因果模型按照功能依赖来编码未观察到的机制，从而指导因果查询。然而，将机制建模为依赖关系受到了研究所谓新机制观点的哲学家的批评，新机制观点是用来分析各种领域中的机制表示的。例如，哲学家卡特赖特（Cartwright）认为，自然科学中的统计依赖关系是在所谓的法则机器（nomological machines）的背景下发现的。规律机器，如钟摆，是由具有特定能力的相互连接的部件组成的非因果装置。其他突出的例子是人工制品，如汽车或机器人。工件被设计为产生协变。这些可以用贝叶斯网络表示，但是为了提供一个因果理解，有必要了解协变是如何由底层的法则机器产生的。珀尔的理论和新机制方法这两个框架都有局限性：珀尔的理论只提供了一个贫乏的机制知识模型，即变量之间的功能依赖关系，而新机制观点夸大了外行人所知道的。此外，两种方法都缺乏将依赖关系知识和机制表示相结合的形式化理论。一个新的理论（MEC-DEP）将发展，解释因果认知的广度和灵活性作为机制和依赖知识之间的相互作用的产物。MEC-DEP将通过在不同内容领域和人群（成人、儿童、非人类灵长类动物）中调查不同的因果模型、机制和任务来进行实验测试。此外，将开发MEC-DEP的计算模型，并分析机制在心理学研究中的作用。新理论将从根本上改变我们对因果关系的许多领域的思考，这些领域可以从对因果推论的规范性和描述性基础的更深入的理解中获益。

Causal cognition endows humans with impressive abilities. It allows us to predict and explain, plan actions, understand confoundings, make educated guesses about transfer situations, think about counterfactual alternatives, and use and engineer innovative devices. No present psychological theory can provide a complete account of all these competencies within an integrated framework. Moreover, partial theories focusing on specific tasks tend to neglect interactions between competencies. Another shortcoming is that most theories ignore or downplay the important role of background knowledge about mechanisms. In the past years, more advanced theories have been proposed in philosophy and computer science that acknowledge some of these deficits. One recent development is Pearl's theory of causation. Pearl has changed his earlier view that causation can be reduced to covariations. He argues now that abstract structural causal models encoding unobserved mechanisms in terms of functional dependencies guide causal queries. However, modeling mechanisms as dependencies has been criticized by philosophers working on the so-called New Mechanism view, which has been developed to analyze mechanism representations in a variety of domains. For example, the philosopher Cartwright has argued that statistical dependencies in the natural sciences are discovered in the context of so-called nomological machines. Nomological machines, such as a pendulum, are non-causal devices that consist of interconnected components with specific capacities. Other salient examples are artifacts, such as cars or robots. Artifacts are designed to produce covariations. These can be represented by a Bayes net, but to provide a causal understanding, it is necessary to understand how the covariations are generated by the underlying nomological machine. Both frameworks, Pearl's theory and the New Mechanism approach, have limitations: Whereas Pearl's theory only provides an impoverished model of mechanism knowledge in terms of functional dependencies between variables, the New Mechanism view exaggerates what laypeople know. Moreover, a formal theory integrates dependency knowledge and mechanism representations is lacking within both approaches. A new theory (MEC-DEP) will be developed that explains the breadth and flexibility of causal cognition as a product of the interaction between mechanism and dependency knowledge. MEC-DEP will be experimentally tested by investigating different causal models, mechanisms, and tasks in various content domains and populations (adults, children, non-human primates). Additionally, computational models of MEC-DEP will be developed and the role of mechanisms in psychological research will be analyzed. The new theory will fundamentally change our thinking about causality in the many areas that can profit from a deeper understanding of the normative and descriptive basis of causal inferences.