Robust Bayesian Inference for Measurement Error Misspecification: The Berkson and Classical Cases

This repository contains all code needed to recreate the results in the paper "Robust Bayesian Inference for Measurement Error Misspecification: The Berkson and Classical Cases" by Charita Dellaporta and Theodoros Damoulas. Our method uses JAX which is compatible with GPU usage. We use the R code (mental_health_study/bcr_sim.R) provided by Harezlak, Ruppert & Wand (2018) here to compare against the BCR method and the R code provided by the SIMEX package. We further use the California school dataset provided by the AER R package here.

The requirements.txt file contains all the requirements.

Reproducing experiments

• The train.py, train_tls.py, train_classical.py and train_cas.py contain the scripts needed to reproduce the non-linear example (with Berkson error), the linear regression (with the TLS loss), the non-linear regression with Classical error and the California School Test experiment respectively.
• The mental_health_study folder contains code to reproduce the mental health study experiment (more details are provided within the folder).
• The gather_results_* notebooks provide the necessary code to process the outputs of the scripts and plot/print results.

References

Harezlak, J., Ruppert, D. and Wand, M.P., 2018. Semiparametric regression with R (Vol. 109). New York: Springer.