princeBART: Principal Stratification with BART

The princeBART package implements principal stratification with Bayesian Additive Regression Trees (BART) for causal inference when treatment uptake is endogenous and identification relies on an instrument that is as-good-as- random conditional on covariates. It targets the average causal effect among compliers (ATE_C, also known as the Local Average Treatment Effect, LATE), as well as conditional complier effects given covariates (CATE_C(x)).

Main Functions

• prince_BART: Fit the principal stratification BART model

• segment_heterogeneity: Summarize heterogeneous effects by segments

• general_BART: Generalize effects to external populations (PATE)

Model

The model applies to settings with an instrument Z, an endogenous treatment W, and an outcome Y. It requires conditional randomization of Z given covariates X, and allows for both one-sided and two-sided noncompliance. Three principal strata are identified:

• Compliers: W(1) = 1, W(0) = 0 — treatment responds to instrument

• Never-takers: W(1) = W(0) = 0 — never treated regardless of Z

• Always-takers: W(1) = W(0) = 1 — always treated regardless of Z

BART is used to flexibly model both the stratum membership probabilities and the potential outcomes within each stratum, allowing for heterogeneous conditional complier effects CATE_C(x) across covariate values.

Author

Maintainer: Lucas Godoy Garraza lgodoygarraz@umass.edu (ORCID)