Bayesian Semiparametric Analysis of Dynamic Treatment Regimes

Summary

This repository contains sample R code for Bayesian semiparametric analysis of dynamic treatment regimes via the Gamma Process models developed here: https://doi.org/10.1093/biostatistics/kxad035

Slightly outdated arxiv version here: https://arxiv.org/abs/2211.16393

Citation

Please use the following .bib citation if you are using LaTeX:

@article{Oganisian2024,
    author = {Oganisian, Arman and Getz, Kelly D and Alonzo, Todd A and Aplenc, Richard and Roy, Jason A},
    title = {Bayesian semiparametric model for sequential treatment decisions with informative timing},
    journal = {Biostatistics},
    volume = {25},
    number = {4},
    pages = {947-961},
    year = {2024},
    month = {01},
    issn = {1465-4644},
    doi = {10.1093/biostatistics/kxad035}
}

File Description and Instructions

The main analysis file is run_bsp_models.R, which takes as input the dataset data/sample_data.Rdata and (1) runs Bayesian semi-parametric models and (2) Bayesian g-computation procedure - both described in the associated paper. See comments in run_bsp_models.R for a detailed description of the data structure and variable names in sample_data.Rdata. Briefly, it contains data on 1,000 subjects who undergo a maximum of four treatment courses along with four baseline covariates and two time-varying covariates.

Part (1) is done by calling the function bayes_dtr_mcmc() contained in the file helper_functions/bayes_dtr_mcmc.R which runs the MCMC algorithm and outputs posterior draws of the model parameters. The function bayes_dtr_mcmc() in turn depends on several helper functions (e.g.such as conditional posteriors) contained in helper_functions/mcmc_helpers.R. bayes_dtr_mcmc() will print progress updates every 1,000 iterations.

Part (2) uses the posterior draws from (1) to evaluate the causal effect of a specified dynamic treatment rule (DTR) via a g-computation procedure. This is done via the gcomp_bayes() function contained in helper_functions/gcomp_bayes.R.

The file run_bsp_models.R runs the gcomp procedure under two specified rules and then plots the marginal survival curve under each. This is contained in output/survivalplots.png:

Optimization can be done by looping gcomp_bayes() across a series of rules.

Runtime

The run-time for bayes_dtr_mcmc() is about 12 minutes for the given data set containing 1,000 subjects with 10,000 total iterations and the following session info. Under the same settings, the run time for gcomp_bayes() is about 1.5 minutes per rule.

R version 4.1.0 (2021-05-18)
Platform: x86_64-apple-darwin17.0 (64-bit)
Running under: macOS Big Sur 11.7.4

Contact

Corresponding Author: Arman Oganisian (email:arman_oganisian@brown.edu)