Introduction
The distinction between general causation (e.g., "smoking causes cancer") and actual causation (e.g., "did this specific cigarette cause this patient's tumor?") is a central challenge in causal inference. Traditional approaches, often rooted in Judea Pearl’s Deterministic Structural Causal Models (DSCMs), frequently rely on subjective intuitions to resolve complex scenarios like "Late Preemption." Sander Beckers proposes a shift toward a functional role for actual causation, utilizing Nondeterministic Structural Causal Models (NSCMs) to provide a more robust and objective framework.
Theoretical Framework: Nondeterministic Models
Unlike DSCMs, which assume perfectly predictable outcomes given all variables ($Y = f(X)$), NSCMs allow for multi-valued equations ($Y \in f(X)$). This nondeterminism captures real-world uncertainty and potential "failures" in causal chains.
As shown in the figure above, the Suzy and Billy example (ST: Suzy Throwing, BT: Billy Throwing, SH: Suzy Hitting, BH: Billy Hitting, BS: Bottle Shattering) can be modeled with specific logical dependencies. In this scenario, Suzy hitting the bottle ($SH$) prevents Billy's hit ($BH$) from occurring, even if he throws ($BT$).
The Logic of Actual Causation
Beckers' framework builds on several key logical operations to redefine how we identify causes:
- Actualized Refinement: This operation "locks in" observed behaviors within a nondeterministic model, focusing on the specific outcome that occurred in a single instance.
- Interventional vs. Observational Dependence: In NSCMs, the timing of an observation relative to an intervention matters. This distinction is crucial for causal discovery—the process of identifying the underlying structure of the world.
- Structural Simplification: This is the core mechanism for identifying causes. The framework seeks a "simpler" version of the causal model (one with fewer edges) that remains consistent with observations. An event $X$ is an actual cause of $Y$ if $Y$ counterfactually depends on $X$ within this simplified model.
Key Insights and Results
The shift to nondeterminism increases the expressivity of causal models, allowing them to align more closely with probabilistic models like the do-calculus. By viewing causation as a tool for communication and abduction (inference to the best explanation), Beckers provides a unified definition of actual causation that handles classic puzzles and edge cases—such as scenarios where an actor is not always accurate—more effectively than previous deterministic iterations.
Major results include Theorem 10, which proves that "direct dependence" in this framework is equivalent to an edge existing in the causal graph, providing a logical bedrock for causal discovery. Furthermore, the new definition maintains consistency with previous work while offering improved accuracy in complex edge cases.
Conclusion and Future Work
Beckers' functional account of actual causation moves the field away from subjective brain-teasers toward a logical bedrock for learning causal structures. While currently reliant on idealized assumptions like exhaustivity (observing every possible intervention), future work aims to extend this approach to fully probabilistic frameworks, specifically applying structural simplifications to Causal Bayesian Networks.
Summary updated