Stan and BART for Causal Inference: Estimating Heterogeneous Treatment Effects Using the Power of Stan and the...

Area of Science:

• Statistics
• Machine Learning
• Causal Inference

Background:

• Machine learning models excel at nonlinear response surfaces, improving treatment effect estimation.
• Standard methods struggle with grouped data, failing to model observation dependence.
• Identifying heterogeneous treatment effects remains challenging with complex data structures.

Purpose of the Study:

• Introduce stan4bart, a novel algorithm combining Bayesian Additive Regression Trees (BART) and Stan.
• Address limitations of existing machine learning and multilevel models for grouped data.
• Improve estimation of average, subgroup, and individual-level treatment effects.

Main Methods:

• Developed stan4bart, integrating BART's flexibility with Stan's computational efficiency.
• Applied the algorithm to estimate treatment effects in multilevel data structures.
• Compared stan4bart performance against flexible and multilevel approaches.

Main Results:

• stan4bart demonstrated stronger performance than existing flexible methods ignoring multilevel structure.
• The algorithm outperformed multilevel approaches with strict parametric forms.
• Accurate estimation of average, subgroup, and individual treatment effects was achieved.

Conclusions:

• stan4bart effectively models nonlinearities and dependence in grouped data.
• The proposed method offers improved accuracy for causal inference in complex settings.
• stan4bart provides a flexible and efficient tool for estimating diverse treatment effects.