The Case for Time in Causal DAGs

Alexander Gilbert Reisach, Alberto Suárez, Sebastian Weichwald, Antoine Chambaz

TL;DR

This paper argues that causal directed acyclic graphs (DAGs), a standard tool for causal inference, are fundamentally incomplete without explicitly incorporating time. The authors propose a formalism for temporal causal variables, demonstrating how this resolves ambiguities in current DAGs and provides a principled way to justify the crucial acyclicity assumption.

What Problem Does This Paper Address?

Causal DAGs, while powerful, are typically treated as "nontemporal." This means they represent relationships between variables without explicitly considering when these variables are measured or when the causal effects occur. This abstraction leads to several problems:

  • Ambiguity: It's unclear whether a nontemporal DAG applies to specific time points, a range of time points, or some other temporal configuration. This makes interpreting the DAG and its implications difficult.
  • Justifying Acyclicity: The acyclicity assumption (no causal loops) is fundamental to DAGs. In a temporal context, acyclicity is often naturally satisfied if causes always precede effects. However, without explicit time, it's hard to know why a DAG is acyclic or if the assumption is truly justified. For instance, if a variable represents an aggregation over time, it might appear to have a causal loop with another variable, even if the underlying instantaneous relationships are acyclic.
  • Real-world Relevance: Many real-world phenomena (e.g., economic inflation, disease progression) are inherently temporal, and variables often represent aggregated measurements over time. Nontemporal DAGs struggle to adequately model these complex temporal dynamics.

The paper aims to bridge this gap by making time an explicit, first-class citizen in causal DAGs.

Methods

The core of the paper is the development of a formalism for composite causal variables that explicitly account for time.

  • Single Time Point Variables: This is the simplest case, where a variable represents a quantity at a single, specific time point. This aligns with traditional time-series causal models.
  • Multiple Time Points Variables: This allows variables to refer to quantities measured at one of several possible time points, with the selection of time points potentially being random. This is formalized using stochastic processes and probability distributions over time points.
  • Aggregating Multiple Time Points Variables: This is crucial for real-world applications, where variables often represent averages, differences, or other aggregations of quantities over a period. An aggregation function is used to combine these time-specific measurements into a single composite variable.

Based on this temporal framework, the authors introduce two key concepts for justifying acyclicity:

  • Time-acyclicity: This occurs when the temporal ordering of variables inherently prevents cycles (e.g., if variable A is always measured before variable B, then B cannot cause A).
  • Effect-acyclicity: This is a stronger condition that applies when time-acyclicity doesn't hold, but there are still no causal cycles in the underlying "atomic" causal DAG (which represents instantaneous causal relationships). This requires examining the causal structure at the most granular temporal level.
  • Total Effect-Acyclicity: An even stronger condition, asserting acyclicity regardless of how variables are temporally defined or aggregated.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Time is Not Optional for Causality: The paper strongly argues that the principle of "causes precede effects" inherently ties causality to time, making temporal considerations essential for any causal model, not just time-series models.
  • Nontemporal DAGs Obscure Meaning: The lack of explicit time in standard DAGs leads to fundamental ambiguities in interpreting what relationships are being modeled and why the acyclicity assumption holds.
  • Composite Variables as the Norm: Most real-world variables are not single time-point measurements but rather aggregations or selections from multiple time points. The proposed formalism directly addresses this.
  • Two Flavors of Acyclicity: Distinguishing between time-acyclicity (due to temporal ordering) and effect-acyclicity (due to the underlying causal structure) provides a more nuanced and principled way to justify the acyclicity assumption.
  • Unrolling Cycles: The paper shows how their formalism can help understand how to "unroll" causal cycles by considering finer temporal resolutions, a concept often mentioned but not formally detailed in prior work.

Key Results

The paper's primary contribution is a formal framework for incorporating time into causal DAGs. This framework:

  • Defines composite causal variables that can represent quantities at single, multiple, or aggregated time points.
  • Introduces and formalizes the concepts of time-acyclicity and effect-acyclicity as distinct conditions for ensuring the acyclicity of causal DAGs.
  • Demonstrates through examples how this temporal perspective can resolve ambiguities in interpreting DAGs and justify the acyclicity assumption in various scenarios, including those with potential cycles.
  • Highlights how this temporal view can impact causal discovery, effect estimation, and the overall applicability of DAGs to real-world problems.

Limitations and Future Work

  • Complexity of Total Effect-Acyclicity: The paper notes that while total effect-acyclicity is a strong assumption that simplifies analysis, it can be difficult to justify in practice, especially when dealing with many interleaved time points.
  • Interplay with Machine Learning: The authors suggest that the practical usefulness of causal machine learning might be limited if composite variables with cyclic causal relationships are not properly handled, implying a need for further research in this area.
  • Further Research Directions: The authors propose several avenues for future work, including:
  • Examining how causal effects change with time distance.
  • Investigating which time points might be representative of others.
  • Analyzing the implications for faithfulness violations.
  • Applying the temporal framework to other classes of graphical causal models.
  • Exploring the discovery of causal effects in settings with time-varying causal relationships.

Summary updated

Abstract

We make the case for incorporating time explicitly into the definition of variables in causal directed acyclic graphs (DAGs). Causality requires that causes precede effects in time, meaning that the causal relationships between variables in one time order may not be the same in another. Therefore, any causal model requires temporal qualification; this applies even if the model does not describe a time series of repeated measurements. We formalize a notion of time for causal variables and argue that it resolves existing ambiguity in causal DAGs and is essential to assessing the validity of the acyclicity assumption. If variables are separated in time, their causal relationship is necessarily acyclic. Otherwise, acyclicity depends on the absence of any causal cycles permitted by the time order. We introduce a formal distinction between these two conditions and lay out their respective implications. We outline connections of our contribution with different strands of the broader causality literature and discuss the ramifications of considering time for the interpretation and applicability of DAGs as causal models.
BibTeX
@misc{reisach2025case,
  author = {Alexander Gilbert Reisach and Alberto Suárez and Sebastian Weichwald and Antoine Chambaz},
  title = {{The Case for Time in Causal DAGs}},
  year = {2025},
  eprint = {2501.19311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.19311},
  howpublished = {arXiv},
}
Figure