Comparison of random forest methods for conditional average treatment effect estimation with a continuous treatment.

Area of Science:

• Statistics
• Machine Learning
• Econometrics

Background:

• Estimating conditional average treatment effects (CATE) with continuous treatments and responses presents significant challenges.
• Existing methods often struggle with complex relationships and confounding factors inherent in observational data.

Purpose of the Study:

• To evaluate the effectiveness of random forests for estimating CATE with continuous outcomes and treatments.
• To compare two distinct random forest strategies for CATE estimation.
• To investigate the impact of confounding, colliding effects, and data centering on estimation accuracy.

Main Methods:

• Utilized random forests, exploring two primary approaches: split rules optimizing treatment effect heterogeneity and using a proxy target variable.
• Conducted a comprehensive simulation study examining confounding, colliding effects, and the benefits of locally centering treatment and/or response variables.
• Incorporated both established and novel random forest implementations.

Main Results:

• Both random forest approaches demonstrated viability for CATE estimation.
• Locally centering both the treatment and response variables emerged as the generally optimal strategy for improving estimation accuracy.
• The simulation results provided insights into the performance under various confounding scenarios.

Conclusions:

• Random forests offer a robust framework for estimating conditional average treatment effects in continuous settings.
• Strategic data preprocessing, specifically local centering of variables, is crucial for maximizing the performance of these methods.
• The findings offer practical guidance for researchers applying causal inference techniques in fields like health economics.