Temporal Causal Reasoning with (Non-Recursive) Structural Equation Models

Maksim Gladyshev, Natasha Alechina, Mehdi Dastani, Dragan Doder, Brian LoganAAAI 2025 · DOI

TL;DR

This paper introduces a novel temporal interpretation of Structural Equation Models (SEMs) that allows for causal reasoning about dynamic systems, even those with feedback loops. They propose a new temporal logic, CPLTL, which integrates time-sensitive interventions and past/future temporal operators, and demonstrate its efficient model-checking.

What Problem Does This Paper Address?

Traditional Structural Equation Models (SEMs) are powerful tools for representing causal relationships between variables, but they are primarily static. This means they struggle to capture the temporal dynamics of real-world phenomena where the timing and duration of events are crucial for determining causality. For instance, a medical treatment's effectiveness might depend not just on whether it was administered, but when and for how long. Existing work on temporal reasoning often uses logics like LTL, but integrating them with the causal insights from SEMs, especially for "actual causality" (determining if a specific event was a cause of another), has been challenging. Furthermore, many SEM-based causality approaches are limited to "recursive" models (no cycles), which restricts their applicability to systems with feedback loops.

Methods

The authors propose a new interpretation of SEMs where they are viewed as mechanisms that transform dynamics of exogenous variables into dynamics of endogenous variables over time. This is achieved by:

  • Temporal Contexts: Instead of a static assignment of exogenous variables, they use a "temporal context" – an infinite sequence of assignments representing how exogenous variables change over time.
  • Computations: A causal model, a temporal context, and an initial configuration of endogenous variables generate a "computation," which is an infinite sequence of assignments to endogenous variables over time.
  • Time-Sensitive Interventions: Interventions are made time-sensitive, denoted as Y(n) ← y, meaning variable Y is set to value y at time step n.
  • Causal LTL with Past (CPLTL): They introduce a new logic that combines Linear-time Temporal Logic (LTL) with past and future operators, and time-sensitive interventions. This allows expressing statements about past, present, and future causal relationships.
  • Handling Non-Recursive Models: Their temporal interpretation naturally handles non-recursive models (with cycles) by allowing them to represent feedback loops, which are common in real-world systems.
  • Temporal Equivalence: They define notions of temporal equivalence for causal models, allowing for comparison of models that might represent the same underlying process but with different time scales or variable sets.
  • Efficient Model-Checking: They show that CPLTL model-checking can be performed in polynomial time.

Datasets

Not discussed in the paper. The paper focuses on theoretical framework development and uses illustrative examples rather than empirical datasets.

Key Insights and Ideas

  • SEMs as Dynamic Transformers: The core insight is reframing SEMs not just as static mappings but as mechanisms that process temporal sequences, enabling causal reasoning about dynamic systems.
  • Temporal Interpretation Unlocks Non-Recursiveness: The temporal approach elegantly handles non-recursive models, which are often necessary for representing feedback loops, without the technical difficulties encountered in static settings.
  • Time-Sensitive Interventions are Key: Explicitly incorporating the timing of interventions is crucial for accurate temporal causal reasoning, moving beyond static "what-if" scenarios.
  • CPLTL Bridges Causality and Temporal Logic: The proposed logic CPLTL provides a formal language to express complex temporal causal statements, integrating the strengths of both SEMs and temporal logics.
  • Equivalence for Temporal Models: Developing notions of temporal equivalence is vital for understanding when different models accurately capture the same temporal causal behavior, especially when dealing with varying time scales.

Key Results

  • A formal framework for temporal causal reasoning using SEMs is presented.
  • The logic CPLTL is introduced, capable of expressing temporal causal statements with time-sensitive interventions.
  • The framework naturally accommodates non-recursive causal models.
  • The model-checking problem for CPLTL is shown to be solvable in polynomial time.
  • The paper demonstrates that arbitrary-lagged Temporal Structural Equation Models (TSEMs) are equivalent to 1-step lagged TSEMs, simplifying representation.

Limitations and Future Work

  • Finite Models: The current framework primarily discusses finite models. Extending it to models with infinitely many variables or infinite ranges is suggested as future work.
  • Probabilistic Settings: The current work focuses on deterministic settings. Adapting the framework to probabilistic scenarios, including nondeterministic structural equations, is a promising direction.
  • Temporal Equivalence Notions: While temporal equivalence is introduced, the authors suggest exploring more general or weaker notions of equivalence, such as (bi)simulation relations, for temporal causal models.
  • Empirical Validation: The paper is theoretical; future work could involve applying and validating the framework on real-world temporal datasets.

Summary updated

BibTeX
@inproceedings{gladyshev2025temporal,
  author = {Maksim Gladyshev and Natasha Alechina and Mehdi Dastani and Dragan Doder and Brian Logan},
  title = {{Temporal Causal Reasoning with (Non-Recursive) Structural Equation Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i14.33639},
}
Figure