Toward identifiability of total effects in summary causal graphs with latent confounders: an extension of the front-door criterion

Charles K. AssaadarXiv 2024 · arXiv · OpenReview

TL;DR

This paper tackles the challenge of identifying causal "total effects" (the overall impact of an intervention) from observational data when dealing with complex, dynamic systems where not all causal relationships are known or can be precisely timed. The authors extend the well-known "front-door criterion" to "summary causal graphs" (SCGs), which are a more flexible way to represent these systems, allowing for cycles and latent confounders, and provide graphical conditions and a formula for identifying these effects.

What Problem Does This Paper Address?

Estimating the true impact of an intervention (a "total causal effect") is crucial in many fields, from medicine to economics. While experiments are the gold standard, they are often infeasible due to cost or ethical concerns. Researchers then turn to observational data and causal graphs to infer these effects. The "front-door criterion" is a powerful tool for identifying total effects even when unobserved factors (latent confounders) are present, but it typically requires a fully specified, acyclic causal graph.

However, real-world systems are rarely so neat. Causal relationships can evolve over time, and it's often difficult to precisely define the temporal lag between causes and effects. Moreover, fully mapping out all causal links is a daunting task. This paper addresses this gap by focusing on "summary causal graphs" (SCGs), which represent time series data without specifying exact temporal lags and can include cycles. The challenge is to adapt causal identification methods, like the front-door criterion, to these more realistic, partially specified SCGs, especially when latent confounders are present.

Methods

The core of this paper is extending the front-door criterion to SCGs. The authors first define a "Discrete-time dynamic structural causal model" (DTDSCM) and its corresponding "Full-Time Acyclic Directed Mixed Graph" (FT-ADMG). They then introduce the "Summary Causal Graph" (SCG) as a coarser representation of the FT-ADMG, where each node represents a time series.

The key contribution is the development of the "SCG-front-door criterion" (Definition 7), which provides graphical conditions for identifying total effects within an SCG. This criterion builds upon the standard front-door conditions but is adapted for the SCG framework, incorporating specific rules for cycles and latent confounders. Building on this criterion, Theorem 1 provides a do-free formula for calculating the total effect, which is a mathematical expression that allows estimation from observed data without needing to simulate interventions. The proofs for these results rely on a series of lemmas that systematically address how to block confounding paths in the more complex FT-ADMGs compatible with the SCG.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • SCGs are a practical compromise: Fully specified causal graphs are often too complex or impossible to construct for dynamic systems. SCGs offer a more manageable representation that captures essential causal relationships without requiring precise temporal lags, making them suitable for real-world applications.
  • The standard front-door criterion falters in SCGs: The paper demonstrates that directly applying the standard front-door criterion to SCGs can lead to incorrect conclusions about identifiability due to the inherent differences in how SCGs represent temporal relationships and potential cycles.
  • A new SCG-specific front-door criterion is needed: The authors develop a novel set of graphical conditions tailored to SCGs that accounts for cycles and latent confounders, enabling more robust causal inference.
  • Identifiability is possible even with cycles and hidden confounders: The SCG-front-door criterion and the derived Theorem 1 show that total effects can still be identified and estimated in complex scenarios where traditional methods fail.

Key Results

The paper's main results are:

  • Definition 7 (SCG-front-door criterion): A set of graphical conditions sufficient for identifying total effects in SCGs with latent confounding and cycles.
  • Theorem 1: A do-free formula for estimating total effects when the SCG-front-door criterion is met. This formula decomposes the total effect into a series of conditional probabilities that can be estimated from observational data.
  • Demonstration of non-identifiability: The paper also illustrates scenarios (using SCGs in Figure 4) where total effects are not identifiable, highlighting the importance of the proposed conditions.

Limitations and Future Work

The authors acknowledge that their result, while significant, is not complete. The identifiability conditions provided are sufficient but not necessarily necessary. Future work aims to develop a complete identifiability result that covers all cases, along with a corresponding do-free formula that doesn't require information about the underlying distribution. They also suggest relaxing the strict positivity assumption used in Theorem 1.

Summary updated

Abstract

Conducting experiments to estimate total effects can be challenging due to cost, ethical concerns, or practical limitations. As an alternative, researchers often rely on causal graphs to determine if it is possible to identify these effects from observational data. Identifying total effects in fully specified non-temporal causal graphs has garnered considerable attention, with Pearl's front-door criterion enabling the identification of total effects in the presence of latent confounding even when no variable set is sufficient for adjustment. However, specifying a complete causal graph is challenging in many domains. Extending these identifiability results to partially specified graphs is crucial, particularly in dynamic systems where causal relationships evolve over time. This paper addresses the challenge of identifying total effects using a specific and well-known partially specified graph in dynamic systems called a summary causal graph, which does not specify the temporal lag between causal relations and can contain cycles. In particular, this paper presents sufficient graphical conditions for identifying total effects from observational data, even in the presence of hidden confounding and when no variable set is sufficient for adjustment, contributing to the ongoing effort to understand and estimate causal effects from observational data using summary causal graphs.
BibTeX
@misc{assaad2024toward,
  author = {Charles K. Assaad},
  title = {{Toward identifiability of total effects in summary causal graphs with latent confounders: an extension of the front-door criterion}},
  year = {2024},
  eprint = {2406.05805},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.05805},
}
Figure