yongfuyu / intro_bayesian_causal

Repository for Introduction to Bayesian Estimation of Causal Effects

A Practical Introduction to Bayesian Estimation of Causal Effects: Parametric and Nonparametric Approaches

This is the companion GitHub repository for the following working paper (currently under review): https://arxiv.org/abs/2004.07375 .

Please cite the code examples here and discussed in the paper by citing the paper. The BibTex is

@misc{oganisian2020practical,
    title={A Practical Introduction to Bayesian Estimation of Causal Effects: Parametric and Nonparametric Approaches},
    author={Arman Oganisian and Jason A. Roy},
    year={2020},
    eprint={2004.07375},
    archivePrefix={arXiv},
    primaryClass={stat.ME}
}

Software Dependencies

---

All code and analyses generated in R version 3.6.3. We particularly rely on

• Stan to fit GPs and parametric causal models.
• ChiRP (https://stablemarkets.github.io/ChiRPsite/index.html) to fit DP models.
• BayesTree to fit BART models.

Directory

---

• dose_response: contains code implementing model discussed in Section 3.1 of the paper. Code generates Figure 1a.
• partial_pool: contains code implementing model discussed in Section 3.2. Code generates Figure 1b.
• partial_pool: contains code implementing model discussed in Section 3.2. Code generates Figure 1b.
• g_comp: contains code estimating model discussed in Section 4.1 using Ridge prior in Equation (10). Generates Figure 2a.
• sensitivity: contains code implementing sensitivity analysis for ignorability violations (Section 5). Generates Figure 2b.
• Nonparametrics: contains code implementing DP, GP, and BART models. The file npbayes.R generates Figure 3a-c. The filte npbayes_ATE.R uses specified models to estimate average treatment effects (ATEs) and generates Figure 3d.

Causal and Bayesian Topics

---

The paper touches on the following topics

• Standardization (i.e. g-computation in the point-treatment setting).
• G-computation for time-varying treatments.
  • Estimating effects of static regimes.
  • Estimating effects of dynamic regimes.
• Performing sensitivity analyses around causal assumptions via priors.

In terms of Bayesian models we touch upon

• Bayesian bootstrapping.
• Ridge-like and horseshoe priors for sparsity in high-dimensional regressions.
• Hierarchical priors that induce partial pooling of conditional causal effects.
• Dirichlet Process (DP) priors.
• Bayesian Additive Regression Trees (BART).
• Gaussian process (GP) priors.