Abstract
We make the case for incorporating a notion of time into causal directed acyclic graphs (DAGs). We demonstrate that nontemporal causal DAGs are ambiguous and obstruct justification of the acyclicity assumption. Assuming that causes precede effects, causal relationships are relative to the time order, and causal DAGs require temporal qualification. We propose a formalization via composite causal variables that refer to quantities at one or multiple time points. We emphasize that the acyclicity assumption requires different justifications depending on whether the time order allows cycles. We conclude by discussing implications for the interpretation and applicability of DAGs as causal models.
Introduction: The Ambiguity of Nontemporal DAGs
In the field of causal inference, Directed Acyclic Graphs (DAGs) are the standard tool for mapping out causal relationships. Traditionally, these DAGs are treated as nontemporal, modeling generic variables (e.g., $A \to H$, or Aspirin causes Headache relief) without explicitly defining when these events occur.
However, causality is strictly bound by time: causes must precede effects. Ignoring time creates severe ambiguities. For instance, if a DAG shows $A \to H$, the lack of a reverse arrow might mean that a headache never causes one to take aspirin, or it might just mean that a future headache doesn't cause past aspirin intake. Without a notion of time, it is difficult to rigorously justify the "acyclicity" assumption—the core requirement that there are no feedback loops. In the real world, feedback loops are common (e.g., headaches cause aspirin intake, which relieves headaches), and cycles only disappear when we precisely define the timeframes of our variables.
Proposed Framework: Composite Causal Variables
To resolve this ambiguity, the authors propose a new formalization that injects time directly into the definition of causal variables. Instead of generic variables, they introduce composite causal variables. A composite variable is built on top of an underlying "atomic causal DAG" where every variable is a single measurement at a single, strict point in time.
A composite variable can represent:
- A single time point: e.g., Temperature at exactly 12:00 PM.
- A mixture of time points: e.g., A medical measurement taken at a random time depending on doctor availability.
- An aggregation of time points: e.g., "Smoking habit," which aggregates thousands of individual actions over years.
By formalizing variables this way, the authors mathematically separate the concept of acyclicity into two distinct phenomena:
- Time-acyclicity: Acyclicity guaranteed simply by the calendar. Variable A is strictly measured before Variable B, so B cannot possibly cause A.
- Effect-acyclicity: The time windows of Variable A and Variable B overlap (meaning the time order allows for a cycle), but the underlying physical mechanism simply doesn't contain a feedback loop.
Key Insights and Ideas
- Cycles are a symptom of temporal aggregation: A true causal cycle between two variables requires that at least one of them refers to quantities at multiple time points. If every variable is a single, distinct snapshot in time, acyclicity is automatically guaranteed by the laws of physics.
- Nontemporal DAGs hide strong assumptions: When researchers draw a generic DAG without time, they are often implicitly assuming total effect-acyclicity—meaning no causal cycles exist regardless of how you aggregate the time points. This is a massive, often unrealistic assumption to make without domain-specific proof.
- Faithfulness violations from time aggregation: When you aggregate time-series data into composite variables, you can accidentally create statistical dependencies that trick causal discovery algorithms into drawing edges that point backward in time.
Conclusion
As a theoretical paper, the main results are qualitative: the successful mathematical formalization of composite variables and the proofs separating time-acyclicity from effect-acyclicity. By forcing researchers to define the timeframes of their variables, this framework makes DAGs slightly harder to use casually, but much more rigorous and applicable to real-world observational data. Future work will investigate how causal effects change as the "time distance" between variables increases and explore whether certain single time points can act as statistically valid representatives for aggregated composite variables.
Summary updated