Targeting resources efficiently and justifiably by combining causal machine learning and theory

Area of Science:

• Health Economics
• Biostatistics
• Machine Learning

Background:

• Efficient resource allocation requires accurate estimation of heterogeneous treatment effects (HTE).
• Causal machine learning (ML) can estimate HTE from big data but often lacks interpretability.
• Black-box models hinder justifiable decision-making in resource allocation.

Purpose of the Study:

• To improve resource allocation efficiency and decision acceptance by integrating explainable AI with theory-driven models.
• To identify individuals for physical activity incentives to maximize population health benefits.
• To enhance understanding of treatment effects in public health interventions.

Main Methods:

• Leveraged large-scale health survey data and meta-analysis results.
• Trained causal ML ensembles to estimate HTE.
• Employed explainable AI to extract heterogeneity dimensions, sign, and monotonicity of moderators.
• Integrated findings into a theory-driven generalized linear model with qualitative constraint (GLM_QC) method.

Main Results:

• The proposed methodology improved expected health benefits for diabetes by 11% and heart disease by 9% compared to traditional approaches.
• Qualitative constraints prevented counter-intuitive effects and improved achieved benefits through model regularization.
• The approach enhances decision-making by providing a rationale aligned with existing theory.

Conclusions:

• Integrating explainable AI with causal ML and theory-driven models significantly improves health outcome predictions.
• The GLM_QC method offers a robust framework for interpretable and effective resource allocation in public health.
• This approach provides a justifiable and efficient method for optimizing interventions based on HTE.