NCoRE: Neural Counterfactual Representation Learning for Combinations of Treatments

Sonali Parbhoo, Stefan Bauer, Patrick Schwab

Introduction

In fields such as medicine and economics, interventions are rarely applied in isolation. A patient might receive a "cocktail" of multiple drugs, or a government might implement several economic policies simultaneously. Estimating the "counterfactual"—the predicted outcome of a specific combination of treatments for an individual—is a fundamental challenge in causal inference.

Traditional methods often struggle with the "combinatorial explosion" of treatment possibilities. For instance, with just 15 potential treatments, there are over 32,000 possible combinations. Most existing models treat each combination as a unique category, leading to severe data sparsity and an inability to model interactions between treatments.

Combinatorial treatment scenarios in healthcare and communication.

The NCoRE Framework

To address these challenges, the authors propose NCoRE (Neural Counterfactual Relation Estimation), a deep learning framework designed to explicitly model cross-treatment interactions. Unlike "flat" models that use one-hot encodings for every possible combination, NCoRE employs a structured, branched architecture that allows for information sharing across different treatment sets.

Architecture

The NCoRE model consists of three primary components:

  1. Shared Base Layers: These layers process individual unit covariates (e.g., patient age, genetic markers) to learn a general representation of the individual.
  2. Treatment Arms & Interaction Subnetworks: NCoRE utilizes $k$ treatment arms, one for each base treatment. The core innovation lies in the interaction subnetworks, which are recursively defined. These layers modify the base representation based on the presence of other treatments in the set, enabling the model to learn synergistic or antagonistic effects.
  3. Merge Layer: Finally, a merge layer aggregates the outputs from the active treatment arms to produce the final outcome prediction ($y_T$).
The NCoRE architecture, illustrating the flow from covariates and treatments through shared base layers and interaction subnetworks to the final output.

To mitigate treatment assignment bias—where treatments are not assigned randomly—the authors also introduced Balanced NCoRE, which uses a randomized batch-matching algorithm to ensure the model distinguishes between treatment effects and pre-existing individual characteristics.

Experimental Evaluation

The researchers evaluated NCoRE across several benchmarks of increasing complexity:

  • HIV Simulator: A synthetic environment used to test sensitivity to the number of treatments and assignment bias.
  • Europe 1 & 2: Semi-synthetic datasets based on real-world HIV patient records from the EuResist and ARCA databases.
  • CRISPR Three-way Knockout: Real-world experimental data involving over 3,000 combinations of gene knockouts in cancer cells, testing the model's ability to generalize to entirely unseen combinations.

Key Results and Insights

NCoRE consistently outperformed existing baselines, achieving the lowest Root Mean Squared Error (RMSE) across all tested scenarios.

  • Scalability: As the number of potential treatments increases, the performance gap between NCoRE and traditional methods widens. NCoRE is uniquely robust to the "curse of dimensionality" inherent in combinatorial treatment spaces.
  • Generalization: In the CRISPR knockout experiments, NCoRE successfully predicted the effects of gene combinations that were not present in the training data. This demonstrates its ability to extrapolate by learning the underlying relationships between individual treatments.
  • Efficiency: NCoRE requires significantly fewer samples to reach high accuracy compared to models like GANITE or TARNET, making it particularly valuable for medical domains where data for rare combinations is scarce.

Conclusion

NCoRE represents a significant advancement in estimating counterfactuals for combinatorial treatments. By leveraging a structured neural architecture to model interactions, it overcomes the limitations of data sparsity and interaction blindness. While the current model assumes "unconfoundedness" and focuses on static treatments, future work may extend NCoRE to handle hidden confounders and sequential, longitudinal treatment data.

Summary updated

Abstract

Estimating an individual's potential response to interventions from observational data is of high practical relevance for many domains, such as healthcare, public policy or economics. In this setting, it is often the case that combinations of interventions may be applied simultaneously, for example, multiple prescriptions in healthcare or different fiscal and monetary measures in economics. However, existing methods for counterfactual inference are limited to settings in which actions are not used simultaneously. Here, we present Neural Counterfactual Relation Estimation (NCoRE), a new method for learning counterfactual representations in the combination treatment setting that explicitly models cross-treatment interactions. NCoRE is based on a novel branched conditional neural representation that includes learnt treatment interaction modulators to infer the potential causal generative process underlying the combination of multiple treatments. Our experiments show that NCoRE significantly outperforms existing state-of-the-art methods for counterfactual treatment effect estimation that do not account for the effects of combining multiple treatments across several synthetic, semi-synthetic and real-world benchmarks.
BibTeX
@misc{parbhoo2021ncore,
  author = {Sonali Parbhoo and Stefan Bauer and Patrick Schwab},
  title = {{NCoRE: Neural Counterfactual Representation Learning for Combinations of Treatments}},
  year = {2021},
  eprint = {2103.11175},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.11175},
}
Figure