Disentangling Causal Effects from Sets of Interventions in the Presence of Unobserved Confounders

Olivier Jeunen, Ciarán M. Gilligan-Lee, Rishabh Mehrotra, Mounia LalmasNeurIPS 2022

Introduction

In the real world, we rarely change just one thing at a time. Doctors prescribe drug cocktails, and marketers run multiple concurrent ad campaigns. Standard causal inference often assumes we can observe the effect of a single intervention in isolation. However, when interventions usually come in sets, determining which specific component drove the outcome becomes a complex challenge. If a patient takes three different medications and gets better, which one actually worked? Did they interact?

This problem is significantly exacerbated by the presence of unobserved confounders—hidden variables, such as a patient's lifestyle or genetic predisposition, that influence both the treatments they receive and their eventual outcome.

This paper tackles a specific "data fusion" challenge: Can we mathematically disentangle and learn the individual causal effect of a single intervention ($do(X_i)$) if our training data only consists of observational records and "joint" interventions where multiple variables were changed at once ($do(X_i, X_j)$)?

Methods: Structural Causal Models and Symmetric Structure

The researchers utilize the framework of Structural Causal Models (SCMs) to address this problem. Their approach relies on several key technical pillars:

1. Symmetric Structure The authors focus on "symmetric" causal graphs. In these scenarios, treatments ($X$) are influenced by observed covariates ($C$) and hidden variables ($U$), but do not directly cause each other. This reflects many real-world scenarios where treatments are assigned based on a patient's state rather than one drug causing the prescription of another.

Symmetric Causal Structure

2. Identifiability Proofs The authors first prove that disentangling these effects is impossible for general, unrestricted causal models. However, they demonstrate that it is possible if the system follows a Non-linear Additive Noise Model (ANM) with multivariate Gaussian noise. Essentially, if the "hidden" influence is additive and follows a bell-curve distribution, the underlying math allows for the true causal parameters to be triangulated.

3. The Algorithm To solve the model, they propose an iterative Expectation-Maximization (EM) style algorithm:

  • Step A: Fix the noise covariance and optimize the structural parameters of the model using maximum likelihood.
  • Step B: Fix the parameters and estimate the noise covariance from the residuals (the difference between predicted and actual values).

4. Heterogeneous Effects Unlike simpler models, their approach incorporates observed covariates ($C$), allowing the model to learn how treatment effects vary across different individuals (e.g., a drug might work for seniors but not for teenagers).

Key Insights and Results

The paper highlights that the specific structure of the noise (additive Gaussian) is what allows us to "see through" the confounding. By pooling data from different "regimes" (some observational, some interventional), the model can successfully isolate the true causal parameters.

The authors validated their approach using two primary sources:

  • Synthetic Data: They generated data using complex polynomial functions with second-order interactions to simulate highly non-linear treatment effects. The proposed method achieved near-optimal Mean Absolute Error (MAE) across 12 different "unseen" intervention combinations. A major takeaway is that if you learn the underlying causal mechanism of individual components, you can accurately predict the outcome of a combination of treatments you have never tested before.
  • Semi-synthetic Medical Data: Using the International Stroke Trial (IST) database (19,435 patients), they focused on aspirin and heparin treatments. They injected controlled levels of unobserved confounding into the real-world data. As the strength of the hidden confounding increased, the error for standard machine learning (regression) exploded because it misattributed the effect of hidden variables to the treatments. In contrast, the authors' method remained stable and close to a theoretical "Oracle" model with perfect information.

Furthermore, the model proved to be "consistent," meaning that as it is provided with more data, its estimates reliably converge toward the true causal effect, even in the presence of complex interactions.

Limitations and Future Work

While powerful, the framework has a few notable limitations that pave the way for future research:

  • The Gaussian Assumption: The identifiability proofs rely heavily on the noise being Gaussian. Real-world noise can sometimes be "heavy-tailed" or skewed, which might break the current model's guarantees.
  • No Treatment-to-Treatment Causality: The "Symmetric" assumption precludes cases where one treatment directly causes another (e.g., Drug A causes a side effect that requires Drug B). When treatments directly influence one another, individual causal effects often lose their general identifiability.
Identifiability Challenges with Interacting Treatments
  • Independence of Covariates: The model assumes that observed covariates ($C$) are independent of the hidden noise ($U$). Relaxing this to allow for "confounded covariates" is a suggested next step for expanding the framework's real-world applicability.

Summary updated

BibTeX
@inproceedings{jeunen2022disentangling,
  author = {Olivier Jeunen and Ciarán M. Gilligan-Lee and Rishabh Mehrotra and Mounia Lalmas},
  title = {{Disentangling Causal Effects from Sets of Interventions in the Presence of Unobserved Confounders}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
}
Figure