Semiparametric Causal Sufficient Dimension Reduction of Multidimensional Treatments

Razieh Nabi, Todd McNutt, Ilya Shpitser

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.

Introduction

In classical causal inference, researchers typically address simple binary or univariate treatments. However, real-world medical interventions, such as radiation therapy for cancer, are often high-dimensional. In radiation oncology, the "treatment" is a complex 3D voxel map of doses delivered to various anatomical structures. To analyze such data, dimension reduction is necessary, but standard techniques often fall into two traps: either they reduce dimensions based solely on treatment variance (like PCA), ignoring the outcome, or they ignore confounding factors—baseline factors like age or cancer stage that influence both treatment choice and outcome—leading to biased, non-causal conclusions.

This paper introduces a novel framework for Causal Sufficient Dimension Reduction (SDR), enabling the compression of complex treatments into lower-dimensional spaces while preserving the integrity of causal effect estimation.

Methodology

The proposed approach is rooted in Marginal Structural Models (MSMs). The goal is to identify a reduction function $g(a; \beta)$ such that the counterfactual outcome $E[Y(a)]$ depends on the high-dimensional treatment $a$ only through its lower-dimensional representation.

The authors develop two primary estimators:

  1. Inverse Probability Weighted (IPW) Estimator: Accounts for confounding by weighting observations based on their propensity for treatment.
  2. Augmented IPW (AIPW) Estimator: A "doubly robust" estimator that remains consistent if either the treatment assignment model or the outcome model is correctly specified.

Additionally, the framework includes a bootstrap-based resampling procedure to objectively determine the optimal number of dimensions ($d$) for the reduction, moving beyond heuristic selection.

Results and Evaluation

The framework was validated using both synthetic simulations and a real-world radiation oncology dataset from Johns Hopkins Hospital involving 613 patients.

Simulation Performance

In controlled simulations, the AIPW estimator demonstrated superior accuracy compared to PCA, standard regression-based SDR, and basic IPW. PCA's failure stems from its outcome-agnostic nature, while standard SDR produces biased results by ignoring the mechanisms of treatment assignment (confounding).

Performance comparison across two simulation cases. The AIPW estimator consistently achieves lower error/bias compared to Regression (Reg), IPW, and PCA methods.

Visualizing the Causal Surface

A critical advantage of this method is its ability to preserve the entire causal response surface. Heatmap comparisons reveal that the AIPW method recovers a causal map nearly identical to the ground truth, whereas standard regression-based SDR produces a significantly distorted representation.

Comparison of causal response surfaces: True Effects (left), standard Regression Effects (center), and the proposed AIPW Effects (right). The AIPW estimator successfully recovers the true causal structure despite confounding.

Clinical Application

In the radiation oncology study, the method successfully reduced 5-dimensional dose-volume histograms into a single dimension. The results confirmed that increasing radiation doses to the parotid glands causally increases the severity of post-treatment weight loss, providing a mathematically rigorous foundation for clinical intuition.

Conclusion and Future Work

This work bridges the gap between dimension reduction and causal inference for multidimensional treatments. While highly effective, the authors note that extremely high-dimensional confounders may require integration with machine learning tools like Double Machine Learning. Future research will focus on "voxel-based" analysis—treating entire 3D images as treatments—and extending the framework to longitudinal settings where treatments are adjusted 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