Meeting Summary: All scientific models abstract away from the details of their target systems. Probabilistic causal models, such as those developed by Pearl and Spirtes et al., are no exception. Scientists have to make choices about the level of detail with which causal models should be defined, and precisely which details should be included in those models. Formalizing these choices in a precise way allows us to bring precision to classic philosophical debates as to whether higher-level models are reducible to more fundamental models, while also addressing a host of other questions about the relationship between higher and lower levels of description. This working group is organized around two key questions: 1. What logic governs our choices regarding the optimal level of abstraction for a causal model? 2. How does the relationship between more abstract and more fundamental causal models shed light on the nature of abstraction and idealization in the sciences?
In discussing the nature of idealization and abstraction in causal modeling, all participants in this working group rely on either informal or formal concepts from information theory, as developed by Shannon. Concepts from information theory allow us to draw connections between abstraction in causal modeling and data compression in the information sciences, which we believe to be a fruitful methodology for addressing the questions listed above.