Joint Causal Inference from Multiple Contexts

Joris M. Mooij, Sara Magliacane, Tom ClaassenJMLR 2020 · OpenReview

TL;DR

This paper introduces Joint Causal Inference (JCI), a unified framework for discovering causal relationships from data collected across multiple contexts (e.g., different experimental conditions or interventions). JCI elegantly combines insights from both experimental and observational causal discovery, enabling more robust and accurate causal inference, especially in complex systems.

What Problem Does This Paper Address?

The gold standard for establishing causal relationships is experimentation, but it's often impractical, unethical, or too expensive. While purely observational causal discovery methods have advanced, they struggle with complex scenarios involving multiple contexts or interventions. Existing methods often treat these contexts separately or require detailed knowledge about intervention types and targets, limiting their applicability. This paper addresses the challenge of inferring causal relationships from diverse datasets collected under various conditions, aiming to unify experimental and observational approaches into a single, flexible framework.

Methods

The core contribution is the Joint Causal Inference (JCI) framework. JCI treats system variables and context variables (which describe the experimental conditions or interventions) as endogenous variables within a single, unified Structural Causal Model (SCM). This "internalization" of context allows existing causal discovery algorithms to be applied to the pooled data, leveraging background knowledge about the relationships between context and system variables. JCI is a framework, not a specific algorithm, and can be implemented with various causal discovery methods. The paper also proposes adaptations of existing algorithms like FCI (FCI-JCI) and ASD (ASD-JCI) to work within the JCI framework, enhancing their capabilities.

Datasets

The paper evaluates its methods on:

  • Synthetic data: Generated from acyclic and cyclic Structural Causal Models with various types of interventions (mechanism changes and perfect interventions).
  • Real-world data: Flow cytometry protein expression data from Sachs et al. (2005), a benchmark dataset in causal discovery.

No new datasets were introduced.

Key Insights and Ideas

  • Unifying Experimentation and Observation: JCI bridges the gap between traditional experimental causal discovery (like Randomized Controlled Trials) and purely observational methods by treating contexts as part of a larger causal model.
  • Context as Endogenous Variables: By modeling context variables as endogenous within a larger "meta-system," JCI allows for a unified treatment of diverse experimental setups and interventions.
  • Flexibility in Intervention Modeling: JCI can handle various intervention types (perfect, imperfect, stochastic) without requiring prior knowledge of their targets or exact mechanisms, making it applicable to real-world scenarios where such information is often unknown.
  • Leveraging Background Knowledge: The framework allows for incorporating specific assumptions about the relationships between context and system variables (JCI Assumptions 1-3), which can significantly improve accuracy and efficiency.
  • Pooling Data is Key: A crucial insight is that pooling data from different contexts and analyzing it jointly, rather than in isolation, unlocks powerful causal signals that would otherwise be missed.

Key Results

  • Outperforms State-of-the-Art: JCI implementations, particularly ASD-JCI and FCI-JCI variants, consistently outperformed baseline methods on synthetic data across various tasks (predicting ancestral relations, direct causal relations, confounders, and intervention targets).
  • Significant Accuracy Gains: JCI methods achieved near-perfect precision in discovering ancestral causal relations, a task where purely observational methods performed poorly.
  • Robustness to Intervention Types: JCI methods performed well even with imperfect interventions (mechanism changes), unlike some methods that assume perfect interventions.
  • Flow Cytometry Data: Applying FCI-JCI variants to the Sachs et al. (2005) flow cytometry data yielded promising results, suggesting the framework's utility for complex biological systems, despite challenges in ground truth validation.
  • Computational Trade-offs: While ASD-JCI is highly accurate, its computational cost increases significantly with the number of variables. FCI-JCI offers a faster alternative with competitive accuracy.

Limitations and Future Work

  • Faithfulness Assumption: The framework relies on a faithfulness assumption, which might be violated in certain real-world scenarios, especially with deterministic relationships or specific experimental designs. The paper discusses workarounds but acknowledges this as a potential limitation.
  • Computational Cost: ASD-JCI can become computationally expensive for larger models.
  • Identifiability: While JCI improves identifiability, some causal relationships may still remain unidentifiable, especially in cyclic or highly complex settings.
  • Future Work: The authors suggest several directions:
  • Weakening the faithfulness assumption for more general experimental designs.
  • Handling datasets with non-identical but overlapping sets of observed variables.
  • Developing algorithms that require less computation time.
  • Addressing selection bias more comprehensively.
  • Applying JCI to more real-world applications.
  • Extending JCI-based causal discovery with causal prediction techniques.

Summary updated

Abstract

The gold standard for discovering causal relations is by means of experimentation. Over the last decades, alternative methods have been proposed that can infer causal relations between variables from certain statistical patterns in purely observational data. We introduce Joint Causal Inference (JCI), a novel approach to causal discovery from multiple data sets from different contexts that elegantly unifies both approaches. JCI is a causal modeling framework rather than a specific algorithm, and it can be implemented using any causal discovery algorithm that can take into account certain background knowledge. JCI can deal with different types of interventions (e.g., perfect, imperfect, stochastic, etc.) in a unified fashion, and does not require knowledge of intervention targets or types in case of interventional data. We explain how several well-known causal discovery algorithms can be seen as addressing special cases of the JCI framework, and we also propose novel implementations that extend existing causal discovery methods for purely observational data to the JCI setting. We evaluate different JCI implementations on synthetic data and on flow cytometry protein expression data and conclude that JCI implementations can considerably outperform state-of-the-art causal discovery algorithms.
BibTeX
@article{mooij2020joint,
  author = {Joris M. Mooij and Sara Magliacane and Tom Claassen},
  title = {{Joint Causal Inference from Multiple Contexts}},
  journal = {J. Mach. Learn. Res.},
  year = {2020},
  url = {https://openreview.net/forum?id=9umPwM3VZa},
}
Figure