Robust CATE estimation using novel ensemble methods

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Estimating Conditional Average Treatment Effects (CATE) is vital for analyzing treatment effect heterogeneity in clinical trials.
• Existing CATE estimation methods, such as causal forests and meta-learners, exhibit limitations and struggle in various scenarios.

Purpose of the Study:

• To develop robust ensemble methods for CATE estimation that enhance prediction stability and performance.
• To address the limitations of current CATE estimators in real-world, uncertain data-generating processes.

Main Methods:

• Proposed two novel ensemble methods: Stacked X-Learner (using X-Learner with model stacking) and Consensus Based Averaging (CBA).
• Evaluated method performance across diverse scenarios varying in complexity, sample size, and underlying mechanisms, including a PD-L1 inhibition pathway model.

Main Results:

• Both proposed ensemble methods demonstrated strong performance across a wide range of tested scenarios.
• The Stacked X-Learner showed superior performance compared to other ensemble methods like R-Stacking and Causal-Stacking.

Conclusions:

• The Stacked X-Learner and CBA offer more reliable and stable CATE estimation than existing methods.
• These ensemble approaches are effective in diverse clinical trial settings, improving the understanding of treatment effect heterogeneity.