Recursive partitioning for heterogeneous causal effects

Area of Science:

• Causal inference
• Statistical modeling
• Machine learning

Background:

• Estimating causal effects is crucial in experimental and observational studies.
• Understanding heterogeneity in treatment effects across different population subsets is a key challenge.
• Existing methods may struggle with high dimensionality and require strong assumptions.

Purpose of the Study:

• To propose methods for estimating heterogeneity in causal effects.
• To develop data-driven approaches for partitioning populations based on treatment effects.
• To enable valid confidence intervals for treatment effects, even with many covariates.

Main Methods:

• Utilizing regression tree methods adapted for treatment effect estimation.
• Implementing an "honest" estimation approach using separate data samples for partitioning and effect estimation.
• Developing a model selection criterion that accounts for bias elimination and variance in subpopulations.

Main Results:

• Honest estimation achieved nominal coverage for 90% confidence intervals in simulations.
• Non-honest approaches resulted in significantly lower coverage (74%-84%).
• The cost in mean squared error for honest estimation ranged from 7-22%.

Conclusions:

• The proposed "honest" estimation method enhances the reliability of causal effect heterogeneity analysis.
• This approach provides valid confidence intervals without sparsity assumptions.
• The methods are robust and improve upon standard techniques in simulation studies.