Markov categories, causal theories, and the do-calculus
This provides a foundational, abstract framework for causal inference that is more widely applicable and conceptually clearer, benefiting researchers in statistics and machine learning.
The paper tackles the problem of formalizing causal reasoning over directed acyclic graphs by developing a category-theoretic framework that abstracts away from specific causal models like structural equation models, resulting in a simpler version of Pearl's do-calculus that is shown to be as strong as the full version.
We give a category-theoretic treatment of causal models that formalizes the syntax for causal reasoning over a directed acyclic graph (DAG) by associating a free Markov category with the DAG in a canonical way. This framework enables us to define and study important concepts in causal reasoning from an abstract and "purely causal" point of view, such as causal independence/separation, causal conditionals, and decomposition of intervention effects. Our results regarding these concepts abstract away from the details of the commonly adopted causal models such as (recursive) structural equation models or causal Bayesian networks. They are therefore more widely applicable and in a way conceptually clearer. Our results are also intimately related to Judea Pearl's celebrated do-calculus, and yield a syntactic version of a core part of the calculus that is inherited in all causal models. In particular, it induces a simpler and specialized version of Pearl's do-calculus in the context of causal Bayesian networks, which we show is as strong as the full version.