uprokevin/Bayesian-Causal-Inference

Bayesian Causal Inference

Table of Contents

• Bayesian Causal Inference
  • Table of Contents
  • About The Project
  • Model
    • Built With
  • Usage
  • License
  • Contact
  • Acknowledgements

About The Project

A Bayesian alternative for Difference-in-difference causal model based on multi-task Gaussian Process regression.

Model

Our model follows the classic setup of Diff-in-diff model. For units in both treatment and control group, we obtain time and unit dependent potentially noisy observation Y_it and covariates X_it. We assume for now that there are no interative effects in the data generation process.

$$f(x,t,g) = h(x) + f_g(t) + u(t)+ e$$

where h(x) is the covariate effect, f(t) is the time trend for group g and u(t) is unit-specific trend. We place independent Gaussian Process priors on h and u but a joint GP on [$f_1$, $f_2$], so the outcome induces to another GP.

Built With

• Python
• Pytorch
• Gpytorch

Usage

The script localnews.m shows how to obtain MAP estimator for the multi-task GP model. Check load_data.m and localnewsmodel.m for how to customize the loading of data and specifying mean/covariance/prior function.

License

Distributed under the MIT License. See LICENSE for more information.

Contact

Yehu Chen - chenyehu@wustl.edu

Acknowledgements

• gpytorch
• GitHub Pages