Abstract
We address the estimation of conditional average treatment effects (CATEs) for structured treatments (e.g., graphs, images, texts). Given a weak condition on the effect, we propose the generalized Robinson decomposition, which (i) isolates the causal estimand (reducing regularization bias), (ii) allows one to plug in arbitrary models for learning, and (iii) possesses a quasi-oracle convergence guarantee under mild assumptions. In experiments with small-world and molecular graphs we demonstrate that our approach outperforms prior work in CATE estimation.
Introduction
In classical causal inference, treatments are typically represented as binary indicators (e.g., taking a pill) or scalar values (e.g., dosage). However, many real-world "treatments" are complex structures, such as molecular graphs in drug discovery, nutritional labels in text, or regional policies represented as images. Estimating the Conditional Average Treatment Effect (CATE)—the specific response of an individual to a specific treatment—in these high-dimensional spaces presents significant challenges.
The primary hurdles are twofold. First, we face the fundamental problem of causal inference: we only observe one treatment outcome per individual, leaving the counterfactuals unobserved. Second, confounding variables (covariates) often influence both the treatment assignment and the outcome, creating non-causal associations that can bias estimation. Existing methods for binary or scalar treatments do not scale naturally to the infinite variety of structured data.
Methodology
Generalized Robinson Decomposition (GRD)
The authors bridge this gap by generalizing the Robinson decomposition, a classic statistical tool used to "partial out" covariate effects and isolate causal signals. They extend this to treatments representable as continuous embeddings through the Product Effect Assumption. This assumption posits that the outcome is an inner product of covariate features and treatment features. Crucially, the authors prove this assumption is "universal," meaning it can approximate any continuous function given sufficient dimensionality.
Structured Intervention Networks (SIN)
To implement the GRD in a deep learning framework, the authors developed Structured Intervention Networks (SIN). SIN utilizes a two-stage neural network training algorithm designed to isolate the causal component of the observed signal.
- Stage 1: A "nuisance" model is trained to estimate the mean outcome based solely on covariates.
- Stage 2: The model simultaneously learns feature maps for covariates and treatments, along with "propensity features"—the expected treatment embedding given an individual's history.
By employing alternating gradient descent, SIN prevents the model from over-fitting non-causal associations, thereby reducing regularization bias that often "washes out" treatment effects in standard regression models.
Experimental Evaluation
The framework was evaluated on two primary simulations with known ground-truth causal effects:
- Small-World (SW): A synthetic dataset involving 1,000 units and 200 Watts–Strogatz small-world graphs as treatments.
- TCGA (The Cancer Genomic Atlas): A semi-synthetic dataset pairing real-world gene expression data from ~9,600 cancer patients with 10,000 molecular graphs from the QM9 chemical database.
Key Results
SIN consistently outperformed state-of-the-art methods like GraphITE and standard GNN regressions. In the TCGA molecular graph experiments, SIN achieved a WPEHE@6 (CATE error metric) of 10.98, significantly lower than GraphITE (30.31) and standard GNNs (30.67).
Furthermore, SIN demonstrated remarkable stability as selection bias increased. While standard models degraded significantly when certain patients were more likely to receive specific treatments, SIN successfully "un-confounded" the data. The model also showed strong generalization capabilities, accurately predicting effects for "out-of-sample" molecular structures not seen during training.
Discussion and Conclusion
The proposed framework introduces several key innovations for causal inference with structured data. By moving from propensity scores to propensity features (expected embeddings), SIN handles high-dimensional treatment spaces effectively. Additionally, the paper provides a quasi-oracle guarantee, showing that the final CATE estimator converges at a rate comparable to an "oracle" with ground-truth knowledge, even with imperfectly estimated nuisance functions.
While the method assumes "unconfoundedness" (all relevant variables are measured), it represents a significant step forward in applying causal reasoning to complex, structured domains. Future work may explore the integration of kernel methods for stronger consistency and the use of proxy variables to address unmeasured confounders.
Summary updated