Identifiability of direct effects from summary causal graphs

Simon Ferreira, Charles K. Assaad

Abstract

Dynamic structural causal models (SCMs) are a powerful framework for reasoning in dynamic systems about direct effects which measure how a change in one variable affects another variable while holding all other variables constant. The causal relations in a dynamic structural causal model can be qualitatively represented with a full-time causal graph. Assuming linearity and causal sufficiency and given the full-time causal graph, the direct causal effect is always identifiable and can be estimated from data by adjusting on any set of variables given by the so-called single-door criterion. However, in many application such a graph is not available for various reasons but nevertheless experts have access to an abstraction of the full-time causal graph which represents causal relations between time series while omitting temporal information. This paper presents a complete identifiability result which characterizes all cases for which the direct effect is graphically identifiable from summary causal graphs and gives two sound finite adjustment sets that can be used to estimate the direct effect whenever it is identifiable.

Introduction

In causal inference, a primary goal is to measure the direct effect: how much a change in variable $X$ at one time step directly influences variable $Y$ at a later step, holding all other factors constant. In dynamic systems—such as ecology, economics, or IT monitoring—the underlying process is governed by a Full-Time Causal Graph (FTCG).

However, constructing an FTCG is often practically impossible due to the complexity of temporal interactions. Instead, researchers frequently work with a Summary Causal Graph (SCG), which collapses temporal information into a single snapshot. This abstraction introduces a significant challenge: the emergence of cycles. While standard causal tools like the back-door criterion are designed for Directed Acyclic Graphs (DAGs), SCGs often contain feedback loops (e.g., $X \rightleftarrows Y$). This paper provides the first complete graphical criteria to determine when a direct effect can be uniquely identified from an SCG alone.

Methodology and Graphical Framework

The authors utilize Linear Dynamic Structural Causal Models (SCMs) to derive their results. The core of their methodology involves mapping the relationship between the abstract SCG and the underlying, potentially infinite, FTCG.

Figure 4: Examples of Summary Causal Graphs (a, d) and their compatible Full-Time Causal Graphs (b, c, e, f). The paper characterizes identifiability across all such compatible temporal structures.

Key components of the approach include:

  1. Adapting d-separation: Redefining how paths are "blocked" in graphs that contain cycles to account for the flow of non-causal correlation in dynamic systems.
  2. Theorem of Completeness: Proving an "if and only if" condition (Theorem 1) that characterizes exactly which graphical structures in an SCG render a direct effect non-identifiable.
  3. Adjustment Set Derivation: Proposing two specific adjustment sets ($Z^{Def 11}$ and $Z^{Def 12}$). These sets provide the mathematical "recipe" for selecting variables to control for in a linear regression to consistently estimate the true direct causal link.

Key Insights

  • The Cycle Size Threshold: A critical finding (Corollary 1) states that if a summary graph contains a cycle involving more than two variables (e.g., $X \to Y \to Z \to X$), at least one instantaneous direct effect in that system is guaranteed to be non-identifiable.
  • The Descendant Trap: In static DAGs, adjusting for descendants of the treatment variable is generally avoided to prevent selection bias. However, in dynamic systems with time lags, the authors demonstrate that certain descendants in the SCG may be necessary to block confounding paths in the underlying FTCG.
  • Closing the Identifiability Gap: By providing a complete result, the authors establish the theoretical limits of what can be measured from summary graphs in linear systems without hidden confounders.

Results and Soundness

The paper identifies two specific conditions that break identifiability: cycles on the cause variable $X$ that exclude $Y$, and active non-direct paths where intermediate nodes are descendants of $Y$.

To address identifiable cases, the authors prove the soundness of two adjustment sets:

  • $Z^{Def 11}$: A broad set encompassing ancestors and descendants within a specific time window.
  • $Z^{Def 12}$: A more parsimonious set restricted to ancestors of $Y$, which is often more practical for empirical estimation as it requires fewer variables.

Limitations and Future Directions

The current results rely on the assumptions of linearity and causal sufficiency (no unobserved confounders). Additionally, the underlying FTCG is assumed to be acyclic, meaning no instantaneous feedback occurs at the exact same moment. Future research will focus on extending these proofs to non-parametric (non-linear) models and exploring systems with hidden confounding or cyclic temporal processes.

Summary updated

Abstract

Dynamic structural causal models (SCMs) are a powerful framework for reasoning in dynamic systems about direct effects which measure how a change in one variable affects another variable while holding all other variables constant. The causal relations in a dynamic structural causal model can be qualitatively represented with a full-time causal graph. Assuming linearity and causal sufficiency and given the full-time causal graph, the direct causal effect is always identifiable and can be estimated from data by adjusting on any set of variables given by the so-called single-door criterion. However, in many application such a graph is not available for various reasons but nevertheless experts have access to an abstraction of the full-time causal graph which represents causal relations between time series while omitting temporal information. This paper presents a complete identifiability result which characterizes all cases for which the direct effect is graphically identifiable from summary causal graphs and gives two sound finite adjustment sets that can be used to estimate the direct effect whenever it is identifiable.
BibTeX
@inproceedings{ferreira2024identifiability,
  author = {Simon Ferreira and Charles K. Assaad},
  title = {{Identifiability of direct effects from summary causal graphs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=KqVZrgJ6ds},
}
Figure