Semiparametric Causal Sufficient Dimension Reduction Of Multidimensional Treatments

Razieh Nabi, Todd McNutt, Ilya Shpitser

Abstract

Cause-effect relationships are typically evaluated by comparing the outcome responses to binary treatment values, representing two arms of a hypothetical randomized controlled trial. However, in certain applications, treatments of interest are continuous and high dimensional. For example, understanding the causal relationship between severity of radiation therapy, represented by a high dimensional vector of radiation exposure values and post-treatment side effects is a problem of clinical interest in radiation oncology. An appropriate strategy for making interpretable causal conclusions is to reduce the dimension of treatment. If individual elements of a high dimensional treatment vector weakly affect the outcome, but the overall relationship between the treatment variable and the outcome is strong, careless approaches to dimension reduction may not preserve this relationship. Moreover, methods developed for regression problems do not transfer in a straightforward way to causal inference due to confounding complications between the treatment and outcome. In this paper, we use semiparametric inference theory for structural models to give a general approach to causal sufficient dimension reduction of a high dimensional treatment such that the cause-effect relationship between the treatment and outcome is preserved. We illustrate the utility of our proposal through simulations and a real data application in radiation oncology.

1. Introduction

In classical causal inference, treatments are typically simplified to binary or single continuous variables. However, many real-world applications involve high-dimensional treatments. In radiation oncology, for instance, a patient receives a complex 3D voxel map of radiation doses. Understanding how this treatment causes side effects requires reducing this high-dimensional map into a few interpretable numbers while strictly preserving the causal signal.

Traditional dimension reduction tools like Principal Component Analysis (PCA) or regression-based Sufficient Dimension Reduction (SDR) often fail in this context. PCA ignores the outcome entirely, while standard SDR for regression only identifies associations, which are frequently biased by confounding factors (e.g., patient age or cancer stage) that influence both treatment assignment and outcomes. This paper introduces a novel framework for Causal Sufficient Dimension Reduction (Causal SDR) to address these challenges.

2. Methodology

The goal of Causal SDR is to find a low-dimensional projection of the treatment $A$, denoted as $g(A; \beta)$, such that the counterfactual mean outcome depends on $A$ only through this projection. The authors propose a semiparametric framework using influence functions to derive estimators that avoid restrictive parametric assumptions.

A key contribution is the development of an Augmented Inverse Probability Weighting (AIPW) estimator. This estimator is "2x2 robust," meaning it remains consistent even if some underlying models (such as the treatment assignment or outcome models) are misspecified, provided a specific subset remains correct. The framework also incorporates ideas from Structural Nested Mean Models (SNMMs) to ensure mathematical consistency between complex nuisance models and utilizes a resampling-based procedure to determine the optimal structural dimension for the reduced treatment space.

3. Results

3.1 Simulation Performance

The framework was evaluated using synthetic simulations where the relationship between treatment, confounders, and outcome was non-linear. The AIPW-based Causal SDR was compared against PCA, standard regression SDR (Reg), and simple IPW.

As shown in Figure 1, the AIPW estimator significantly outperformed competing methods in terms of accuracy and stability. While standard regression-based SDR captures associations, it is heavily biased by confounding, leading to higher error rates.

Figure 2 illustrates the "spurious associations" that plague non-causal methods. The regression-based SDR picks up signals at the boundaries that are not present in the ground truth, whereas the AIPW-based Causal SDR accurately recovers the true causal structure.

3.2 Clinical Application: Radiation Oncology

The method was applied to a cohort of 613 head and neck cancer patients. High-dimensional radiation dose maps were successfully reduced to a single dimension ($d=1$). This reduction allowed clinicians to visualize how increasing radiation doses across the parotid glands causally increases post-treatment weight loss, providing a more interpretable tool for treatment planning than raw voxel maps.

4. Conclusion

This research provides a robust framework for reducing the dimensionality of complex treatments without losing the causal signal. By accounting for confounding through a doubly robust semiparametric approach, Causal SDR offers a significant improvement over traditional methods. Future work will explore handling high-dimensional confounders and extending the framework to longitudinal settings where treatments and confounders evolve over time.

Summary updated

Abstract

Cause-effect relationships are typically evaluated by comparing outcome responses to binary treatment values, representing two arms of a hypothetical randomized controlled trial. However, in certain applications, treatments of interest are continuous and multidimensional. For example, understanding the causal relationship between severity of radiation therapy, summarized by a multidimensional vector of radiation exposure values and post-treatment side effects is a problem of clinical interest in radiation oncology. An appropriate strategy for making interpretable causal conclusions is to reduce the dimension of treatment. If individual elements of a multidimensional treatment vector weakly affect the outcome, but the overall relationship between treatment and outcome is strong, careless approaches to dimension reduction may not preserve this relationship. Further, methods developed for regression problems do not directly transfer to causal inference due to confounding complications. In this paper, we use semiparametric inference theory for structural models to give a general approach to causal sufficient dimension reduction of a multidimensional treatment such that the cause-effect relationship between treatment and outcome is preserved. We illustrate the utility of our proposals through simulations and a real data application in radiation oncology.
BibTeX
@inproceedings{nabi2022semiparametric,
  author = {Razieh Nabi and Todd McNutt and Ilya Shpitser},
  title = {{Semiparametric Causal Sufficient Dimension Reduction Of Multidimensional Treatments}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=BFULBwUocxq},
}
Figure