Flexible Machine Learning Estimation of Conditional Average Treatment Effects: A Blessing and a Curse

Area of Science:

• Statistics
• Machine Learning
• Causal Inference

Background:

• Estimating causal effects from observational data relies on assumptions that are difficult to verify.
• Machine learning methods offer potential for studying complex causal effect heterogeneity.
• Existing methods for conditional average treatment effect (ATE) estimation may not capture all individual variations.

Purpose of the Study:

• To investigate the divergence between individual treatment effects and conditional ATE when using causal random forests.
• To develop an extended causal random forest capable of estimating differences in conditional variance between treated and control groups.
• To identify conditions under which individual treatment effect heterogeneity can be quantified.

Main Methods:

• Application of causal random forests to observational data.
• Extension of causal random forests to estimate the difference in conditional variance between treated and control groups.
• Analysis of the relationship between individual treatment effect distribution and conditional ATE distribution.

Main Results:

• Demonstrated that individual treatment effect distributions can differ from conditional ATE distributions.
• Showed that an extended causal random forest can estimate the variance of individual treatment effects when distributions diverge.
• Highlighted the limitations of standard causal random forests in capturing certain types of heterogeneity.

Conclusions:

• Standard causal random forests may fail to accurately estimate individual treatment effect variance when heterogeneity is not fully explained by features.
• An extended causal random forest approach is proposed to better quantify individual treatment effect heterogeneity.
• An additional causal assumption is required to quantify heterogeneity not captured by the conditional ATE distribution.