Machine Learning for Causal Inference

Estimation of causal effects requires making comparisons across groups of observations exposed and not exposed to a treatment or cause. This chapter introduces the building blocks necessary to understand what causal quantities represent conceptually and why they are so difficult to estimate empirically. At a basic level, causal inference methods require fair comparisons. Regression provides one way to condition on confounders in an attempt to create fair comparisons. Boosted Regression Trees emerged as a way to address these issues of overfitting, difficulty in capturing additive structure, and overemphasis on high-level interactions. The mean structure of Bayesian Additive Regression Trees (BART) is the same as the boosted regression tree. The BART approach to causal inference, with its combination of flexible modeling embedded in a Bayesian likelihood framework, provides the opportunity for simultaneous inference on individual-level treatment effects as well as any of a variety of average treatment effects.