Code to accompany Combining chains of Bayesian models with Markov melding

arXiv version

Re-running this code

• The code for the little owls example is in scripts/owsl-example/, and code for the respiratory failure example is in scripts/mimic-example/.
• Re-running the code requires MIMIC-III with all the standard derived concept tables on a Postgres database, running and available at the connection give in these three files.
  • The first time you rerun the analysis you may need to uncomment and run the CREATE OR REPLACE commands at the top of blood-gasses.sql.
• The R package dependencies are managed by renv. Some of the packages are not available on CRAN and are my own forks (the BA template is a fork of rticles that I haven't had the time to submit a PR for), but renv should install the correct versions.
• The analysis is handled by GNU Make, and you will need the latest version (>= 4.3) to run make.
  • Running make with -j 2 or higher will parallelise much of the computation. The MCMC samplers run 5 parallel chains, so ensure you have at least 5 times the number of CPU cores as the requested number of Make jobs.
• I think some parts of the analysis assume certain folders exist but don't necessarily check that they do indeed exist.

Diagnostics

The MCMC diagnostic plots and tables referred to in the text can be found here for the little owls example and here for the MIMIC respiratory failure example. You will need to download the .html files and open them in a browser to view the diagnostics. If you rerun the analyses then the .html files will need to be regenerated from their corresponding .rmd source. This takes about an hour.

Other things

• The logistic regression model mentioned in the discussion is here, and the correspondence from the MIMIC team is here.