Beyond Structural Causal Models: Causal Constraints Models

Tineke Blom, Stephan Bongers, Joris M. MooijUAI 2019 · OpenReview

Abstract

Structural Causal Models (SCMs) provide a popular causal modeling framework. In this work, we show that SCMs are not flexible enough to give a complete causal representation of dynamical systems at equilibrium, especially when the outcome depends on starting conditions. We introduce Causal Constraints Models (CCMs), a generalization that uses flexible constraints to perfectly capture the stationary behavior of complex systems and universal functional laws that SCMs cannot handle.

Introduction

In causal inference, a central goal is predicting how a system behaves at a "steady state" following an intervention. Traditionally, researchers have relied on Structural Causal Models (SCMs), which define each variable as a function of its parents and exogenous noise. However, the authors identify two major flaws in this approach for dynamical systems:

  1. Dependence on Initial Conditions: Many real-world systems, such as chemical reactions, exhibit "semistability," where the final equilibrium depends on the initial state. SCMs are mathematically too rigid to capture this history-dependence.
  2. Functional Laws: Physical laws like the Ideal Gas Law ($PV = nRT$) describe symmetric relationships. Forcing these into an SCM requires an arbitrary causal direction (e.g., $P \to V$), which can lead to logical contradictions under simultaneous interventions.

This paper bridges the gap between complex differential equations and overly simplistic SCMs by proposing Causal Constraints Models (CCMs).

Causal Constraints Models (CCMs)

Unlike SCMs, which assign a specific equation to every variable, a CCM consists of a set of causal constraints. Each constraint is defined by a mathematical relationship and an activation set—the specific intervention scenarios under which that constraint remains valid. This allows the model to represent systems where relationships "break" or change depending on the intervention.

Dynamics of a chemical reaction system showing how equilibrium concentrations depend on initial conditions (semistability).

CCMs are derived from Ordinary Differential Equations (ODEs) by identifying fixed points and incorporating constants of motion (conservation laws). These constants ensure the model retains the "memory" of the initial state that persists at equilibrium.

Case Studies

The authors validate the CCM framework using several classic models:

  • The Basic Enzyme Reaction: Used to prove that SCMs cannot represent initial-condition dependence.
  • The Lotka-Volterra Model: Demonstrates how CCMs handle non-convergent, oscillating systems.
  • The Ideal Gas Law: Shows how CCMs naturally handle symmetric physical constraints.
Equilibrium values ($S^*, C^*, E^*$) for the basic enzyme reaction under various intervention scenarios ($I$).

Key Results

The paper establishes several foundational results:

  • Theorem 1 (Impossibility): Formally proves that no SCM can completely represent the causal semantics of a basic enzyme reaction if it depends on initial states.
  • Theorem 2 (Existence): Proves that for any dynamical system with a unique solution, there always exists a CCM that perfectly describes its equilibrium behavior under any intervention.
  • Parsimony: CCMs can represent complex physical laws more naturally and with fewer equations than equivalent SCMs.

Conclusion and Future Work

CCMs offer a more flexible and accurate representation of equilibrium behavior than SCMs. However, they currently lack the intuitive graphical language (like DAGs) that makes SCMs popular. Future work will focus on developing a graphical representation for CCMs, translating concepts like "d-separation" to the constraint-based framework, and creating algorithms to learn CCMs directly from experimental data.

Summary updated

Abstract

Structural Causal Models (SCMs) provide a popular causal modeling framework. In this work, we show that SCMs are not flexible enough to give a complete causal representation of dynamical systems at...
BibTeX
@inproceedings{blom2019beyond,
  author = {Tineke Blom and Stephan Bongers and Joris M. Mooij},
  title = {{Beyond Structural Causal Models: Causal Constraints Models}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=boDGd9jFbj},
}

Loading related content...

Figure