Core Concepts
Indeterministic causal laws, where the effect does not uniquely determine the cause, require a more general modeling approach than deterministic causal models. The paper proposes a framework of relational causal teams to represent and reason about interventionist counterfactuals in the presence of indeterministic causal laws, and provides complete axiomatizations for this setting.
Abstract
The paper investigates the generalization of causal models to the case of indeterministic causal laws, where the effect does not uniquely determine the cause. It provides an overview of the differences in modeling that this more general perspective enforces, and proposes an implementation of generalized models in the style of causal team semantics.
In indeterministic causal models, the laws are not represented by functions (as in the deterministic case), but more generally by relations. This leads to significant differences in the axiomatization of interventionist counterfactuals compared to the deterministic case. The paper provides strongly complete axiomatizations over the full class of indeterministic models and over its recursive subclass, where cyclic causal relationships are forbidden.
The key insights are:
Uncertainty: Indeterministic laws require a shift from causal models to the more general causal teams, which allow a multiplicity of variable assignments compatible with the causal laws.
Specifying causal laws: In the indeterministic case, the causal laws cannot be simply represented as functions, but need to be specified as relations. This raises challenges in identifying the direct causes of a variable.
Interventions: Interventions on indeterministic models can produce multiple possible scenarios, even in the acyclic (recursive) case. This leads to important differences compared to the deterministic case, such as the failure of composition principles.
Axiomatization: The paper provides a strongly complete axiomatization for the logic of interventionist counterfactuals over the full class of indeterministic models, as well as over the recursive subclass. The axiomatization highlights key differences from the deterministic case, such as the importance of might-counterfactuals and the failure of principles like Composition.