AVERAGE CONTROLLED AND AVERAGE NATURAL MI-CRO DIRECT EFFECTS IN SUMMARY CAUSAL GRAPHS

Simon Ferreira, Charles K. AssaadOpenReview 2026 · Submission to ICLR 2026 · Rejected · OpenReview

TL;DR

This paper provides the first theoretical framework for identifying micro-level direct causal effects (how one specific time point affects another) using only Summary Causal Graphs (SCGs). By establishing specific graphical conditions, the authors demonstrate how to identify these effects even in complex, non-linear systems that involve feedback loops (cycles) and unobserved "hidden" variables.

What Problem Does This Paper Address?

In fields like epidemiology or finance, we often deal with time-series data. Ideally, we would use a "Full-Time" causal graph that maps every variable at every single second. However, these graphs are usually too complex to build. Instead, researchers use Summary Causal Graphs (SCGs), which collapse time: if Smoking causes Lung Cancer at any point, the SCG just shows an arrow from "Smoking" to "Cancer."

The challenge is that SCGs are abstractions. A single arrow in an SCG could represent many different temporal patterns in reality. Furthermore, real-world systems are rarely linear and often contain hidden confounding (unobserved factors affecting both cause and effect) and cycles (feedback loops). While previous work focused on "total effects" or linear models, identifying non-parametric direct effects—the impact of a cause on an effect while holding all other intermediate factors constant—remained an unsolved challenge for these summary representations.

Methods

The authors utilize the framework of Structural Causal Models (SCMs) and Pearl’s do-calculus. Their approach involves:

  1. Defining "Possible Parents": Since an SCG doesn't show specific time lags, the authors define a set of "possible parents" ($PP$) to account for all potential temporal ancestors across all full-time graphs compatible with the summary graph.
  2. Graphical Criteria: They derive sufficient conditions based on the structure of the SCG (e.g., the presence of cycles or bidirected arrows representing hidden confounding).
  3. Identifiability Theorems:
    • Theorem 1 provides conditions for the Average Controlled Micro Direct Effect (CDE).
  4. Theorem 2 provides conditions for the Average Natural Micro Direct Effect (NDE), which is more complex because it involves counterfactual reasoning (what would have happened if a variable took its "natural" value).

Datasets

Not discussed in the paper. This is a theoretical paper focused on mathematical proofs and graphical logic. However, the authors provide a conceptual "Real World Example" in epidemiology, modeling the relationship between emissions regulations, economic activity, air pollution, and respiratory health to illustrate how their theorems apply to complex, cyclic public health data.

Key Insights and Ideas

  • The Ambiguity of Cycles: The authors highlight that cycles in an SCG (e.g., $X \rightleftarrows Y$) are particularly troublesome. A cycle in the summary graph means the underlying temporal graph could have arrows in either direction at different times, making it impossible to "block" certain paths using standard adjustment methods.
  • Direct vs. Total Effects: Direct effects are inherently harder to identify than total effects in non-parametric settings because you must account for all parents of the outcome variable, not just a set of mediators.
  • Necessity of Acyclicity for the Outcome: A major finding is that for a micro direct effect to be identifiable, the outcome variable ($Y$) cannot be part of a cycle with other variables in the SCG.
  • Bridging the Gap: The paper proves that if certain graphical conditions are met in the simplified SCG, a valid sequence of do-calculus rules is guaranteed to exist for any underlying full-time temporal graph.

Key Results

  • CDE Identifiability: The Average Controlled Micro Direct Effect is identifiable if the outcome $Y$ is not in a cycle and there is no hidden confounding between $Y$ and its ancestors. The authors provide a specific adjustment formula to calculate this from observational data.
  • NDE Identifiability: The Average Natural Micro Direct Effect requires stricter conditions, including constraints on the treatment variable ($X$) and ensuring no shared hidden confounders between the ancestors of $X$ and $Y$.
  • Necessary Conditions: The authors prove that their sufficient conditions for CDE identification are also necessary when researchers are limited to "identifiability by adjustment" in settings without hidden variables.

Limitations and Future Work

  • Parent-Centric Focus: The current work focuses on identifying direct effects by controlling for the parents of the effect. The authors note that future work should explore conditions that focus on mediators instead.
  • Beyond Adjustment: The necessity results are currently limited to "identifiability by adjustment." The authors suggest exploring whether these conditions remain necessary for more complex identification strategies (like the front-door criterion).
  • Path-Specific Effects: While this paper covers direct effects, the authors aim to extend this logic to indirect effects and more granular path-specific effects within summary graphs.
  • Empirical Validation: As a theoretical piece, the next logical step is applying these theorems to high-dimensional real-world longitudinal datasets, such as medico-administrative databases.

Summary updated

Abstract

In this paper, we investigate the identifiability of average controlled direct effects and average natural direct effects in causal systems represented by summary causal graphs, which are abstractions of full causal graphs, often used in dynamic systems where cycles and omitted temporal information complicate causal inference. Unlike in the traditional linear setting, where direct effects are typically easier to identify and estimate, non-parametric direct effects, which are crucial for handling real-world complexities, particularly in epidemiological contexts where relationships between variables (e.g., genetic, environmental, and behavioral factors) are often non-linear, are much harder to define and identify. In particular, we give sufficient conditions for identifying average controlled micro direct effect and average natural micro direct effect from summary causal graphs in the presence of hidden confounding. Furthermore, we show that the conditions given for the average controlled micro direct effect become also necessary in the setting where there is no hidden confounding and where we are only interested in identifiability by adjustment.
BibTeX
@misc{ferreira2026average,
  author = {Simon Ferreira and Charles K. Assaad},
  title = {{AVERAGE CONTROLLED AND AVERAGE NATURAL MI-CRO DIRECT EFFECTS IN SUMMARY CAUSAL GRAPHS}},
  year = {2026},
  url = {https://openreview.net/forum?id=hisAy19yMP},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}
Figure