Technical Summary: Formalizing and Falsifying Causal Pathways of Rare Events

1. Problem Statement

The paper addresses the challenge of providing causal explanations for rare events (outliers), such as natural disasters, market crashes, or system failures. While existing root cause analysis methods effectively identify a small set of causal origins, they often fail to capture the specific mechanisms and sub-graphs through which these causes propagate to produce the observed outcome. Furthermore, full causal models are frequently unknown, and standard approaches relying on asymptotic extreme-value theory or continuous variable assumptions are ill-suited for rare but non-extreme events or discrete domains.

The core problem is the lack of a formal, testable notion of a "causal pathway" that bridges simple verbal explanations and detailed causal modeling. Current methods struggle to quantify whether a proposed chain of events constitutes a valid explanation for a specific rare observation, particularly when the full causal graph is not available or when the explanation must remain robust under partial model knowledge.

2. Methodology

2.1 Formalizing Causal Pathways and Explanation Scores

The authors introduce a framework based on Structural Equation Models (SEMs) and Directed Acyclic Graphs (DAGs) to formalize causal pathways for binary events.

• Cluster of Events: A set of binary variables $B$ connected by a causal DAG $C$. A subset $R$ represents root causes where mechanisms are "corrupted" or intervened upon.
• Explanation Score: Building on Oesterle et al. (2025), the paper defines an explanation score $E(x \to y)$ for an event $X=x$ explaining $Y=y$ as:
  $$E(x \to y) := 1 - \frac{\log P(Y=y | do(X=x))}{\log P(Y=y)}$$
  This score measures how much the intervention $do(X=x)$ increases the likelihood of $Y=y$ relative to its prior rarity. A score close to 1 indicates a strong explanation.
• Pathway Explanation Score: To address the limitation that root causes alone do not explain the entire propagation, the authors define a pathway explanation score $E^K_{R \to t}$. This score evaluates whether setting root causes $R$ to 1 renders the entire cluster of events (including mediators) likely, relative to the target event's rarity.
  $$E^K_{R \to t} := 1 - \frac{\log P(B=1 | do(B_R=1))}{\log P(B_t=1)}$$
  This metric penalizes explanations where intermediate events remain unlikely even after the root causes are set, ensuring the pathway is coherent.

2.2 Causal Abstraction and Binarization

To handle continuous or complex variables, the paper proposes Pathway Abstraction, a method to map detailed causal models to simplified binary pathways.

• Feature Monotonicity: Variables are transformed into binary events using feature functions $\tau$ (e.g., $X \ge x$). The framework requires these features to satisfy "feature monotonicity," ensuring that increasing the feature value of a parent increases the probability of the feature value of the child.
• Natural Micro-Realization: To handle the ill-defined nature of interventions on coarse-grained (binary) variables, the authors define a "natural micro-realization" of an intervention $do(B_S=b_S)$. This involves drawing the underlying continuous variables from their conditional distribution given the binary constraint, effectively simulating the intervention in the fine-grained model.
• Accuracy Metric: The quality of an abstraction is measured by the KL-divergence between the interventional distributions of the abstract model and the true projected distribution, normalized by the target event's rarity.

2.3 Evaluation via Consistency Checks

The framework enables falsification of causal explanations through consistency checks:

• Quantitative Consistency: If a pathway is proposed, the probabilities of transitions between nodes (e.g., $P(B_j=1 | B_i=1)$) must be consistent with the explanation score. If the product of these probabilities does not sufficiently increase the likelihood of the target, the explanation is deemed weak.
• Qualitative Consistency: The framework can be used to query agents (e.g., LLMs) about sub-structures of a proposed pathway to verify if the claimed causal links hold under specific contexts.

3. Key Contributions

1. Formal Definition of Causal Pathways: The paper defines a causal pathway not just as a set of root causes, but as a tuple $(C, P, B, B_t, B_R, P_B)$ where $P$ is a subgraph explaining the target. This distinguishes between the full causal graph $C$ (used for context) and the explanatory pathway $P$.
2. Pathway Explanation Score: A novel metric that quantifies the quality of an explanation by assessing whether the root causes render the entire sequence of events in the pathway likely, rather than just the target. This score is additive over disjoint sets of mechanisms and avoids the context-dependency issues of Shapley values.
3. Event-Level Abstraction: The introduction of pathway abstraction allows for the derivation of simple, verbalizable causal explanations from complex, continuous, or partially known models. It provides a rigorous way to handle the "coarse-graining" of variables while maintaining testable implications.
4. Falsifiability: By requiring explicit commitments on which mechanisms behaved as expected and which were altered, the framework renders causal explanations falsifiable. It provides conditions (e.g., Lemma 3.7 on log-likelihood gaps) under which an explanation can be rejected.

4. Results and Demonstrations

• Theoretical Bounds: The authors prove that strong local causal explanations (high pairwise scores) do not guarantee a strong global pathway explanation. They derive necessary conditions for high pathway scores, such as avoiding mediators that are significantly less likely than their parents.
• Relation to Necessity: The paper demonstrates that for rare events, a positive explanation score implies a high probability of necessity (PN), meaning the root cause is likely necessary for the event to occur, even if the score is not perfect.
• LLM Evaluation Case Study: The authors applied their framework to evaluate causal pathways generated by a Large Language Model (LLM) regarding the causes of long-term homelessness.
  • The LLM generated a pathway $A \to B \to C \to D \to E$.
  • By querying the LLM for transition probabilities, the authors calculated a pathway explanation score of $\approx 0.29$.
  • The low score was attributed to a weak link ($C \to D$), where the transition from "lost apartment" to "family estrangement" was not sufficiently probable.
  • This exercise demonstrated that the framework can quantitatively identify weak links in verbal explanations and suggest revisions (e.g., adding a direct link or rephrasing nodes) to improve plausibility.

5. Significance and Claims

The paper claims to advance the field of causal inference and Explainable AI (XAI) by providing a formal, quantitative language for discussing causal explanations of rare events.

• Bridging the Gap: It bridges the gap between simple verbal explanations (which are often intuitive but untestable) and detailed causal modeling (which is often too complex or unknown).
• Falsifiability: The primary significance lies in making causal explanations falsifiable. By requiring that an explanation implies a specific set of interventional probabilities, the framework allows for empirical or logical checks against data or expert belief.
• AI Safety and Transparency: The authors position this work as a step toward assessing the truthfulness of causal statements generated by AI systems. By formalizing explanations as explicit pathways, it becomes possible to detect flawed causal accounts of socially significant events, thereby contributing to the reliability of generative AI.

The authors remain modest, noting that the framework does not automatically prove the validity of a causal DAG or the truth of interventional probabilities; rather, it provides the concepts and metrics necessary to discuss and evaluate plausibility in a data-driven manner.