Individualised Treatment Effects Estimation with Composite Treatments and Composite Outcomes

TL;DR

The paper tackles the challenge of estimating individualised treatment effects when both treatments and outcomes are composite, a setting common in healthcare and social sciences, by introducing H-Learner, a hypernetwork-based framework. H-Learner uses an embedding layer and a task-conditioned hypernetwork to generate instance-specific target learners for each treatment–outcome pair, enabling information sharing across many combinations and mitigating data scarcity. In experiments on synthetic data with varying numbers of treatments and outcomes, H-Learner shows superior or competitive performance relative to state-of-the-art baselines, especially in low-data regimes or when outcomes are numerous. This work broadens causal ML applicability to complex, real-world scenarios with polypharmacy and multi-outcome objectives by providing a scalable, transfer-learning style approach to ITE estimation.

Abstract

Estimating individualised treatment effect (ITE) -- that is the causal effect of a set of variables (also called exposures, treatments, actions, policies, or interventions), referred to as \textit{composite treatments}, on a set of outcome variables of interest, referred to as \textit{composite outcomes}, for a unit from observational data -- remains a fundamental problem in causal inference with applications across disciplines, such as healthcare, economics, education, social science, marketing, and computer science. Previous work in causal machine learning for ITE estimation is limited to simple settings, like single treatments and single outcomes. This hinders their use in complex real-world scenarios; for example, consider studying the effect of different ICU interventions, such as beta-blockers and statins for a patient admitted for heart surgery, on different outcomes of interest such as atrial fibrillation and in-hospital mortality. The limited research into composite treatments and outcomes is primarily due to data scarcity for all treatments and outcomes. To address the above challenges, we propose a novel and innovative hypernetwork-based approach, called \emph{H-Learner}, to solve ITE estimation under composite treatments and composite outcomes, which tackles the data scarcity issue by dynamically sharing information across treatments and outcomes. Our empirical analysis with binary and arbitrary composite treatments and outcomes demonstrates the effectiveness of the proposed approach compared to existing methods.

Figures (4)

• Figure 1: Example: An intensive care unit patient, admitted for heart surgery, needs composite treatments to optimise composite outcomes.
• Figure 2: Architecture of H-Learner: It comprises an embedding layer $e(t,y;\psi)$, a hypernetwork $h(e_{\mathbf{t},\mathbf{y}}; \phi))$, and an ITE learner $f(\mathbf{X};\theta_{t,y})$. The hypernetwork $h$, conditioned on an embedding of treatment-outcome combination $e_{\mathbf{t},\mathbf{y}}$, generates the ITE learner’s $f$ weights $\theta_{t,y}$, enabling dynamic information sharing across treatments and outcomes.
• Figure 3: Comparative study of ITE learners for arbitrary composite treatments and composite outcomes (shaded region represents one standard error over 10 repetitions).
• Figure 4: Comparative study of ITE learners for binary composite treatments and composite outcomes (shaded region represents one standard error over 10 repetitions). H-Learner does not need any assumptions about the interaction among the treatments and can handle composite outcomes, but SCP needs to know the causal structure among treatments and is limited to single outcomes.