BART with Targeted Smoothing: Analysis

This repository contains code to replicate figures and tables in the paper "BART with Targeted Smoothing: An analysis of patient-specific stillbirth risk", by Starling et al (2019).

Note that many of these scripts use the dbarts R package for comparing tsBART with regular BART. This package is available on CRAN (https://cran.r-project.org/package=dbarts).

The following are instructions for the workflow to replicate the analysis.

Dataset.

The raw data is publicly available via the National Center for Health Statistics at the following locations:

• 2005 data: ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/cohortlinkedus/LinkCO05US.zip
• 2006 data: ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/cohortlinkedus/LinkCO06US.zip

We provide several items to support replication of our process:

• A 'data dictionary' for our dataset, which includes a link to the original data source and its documentation.
  • Our data dictionary is located at /data/04_stillbirth-example-data/data-dictionary.pdf.
  • The original dataset and documentation are located at https://www.cdc.gov/nchs/data_access/vitalstatsonline.htm#Period_Linked
• A sample stillbirth dataset, to illustrate the structure and format of our data set post cleaning and processing.
  • The sample dataset is located at data/perinatal-mortality/stillbirth-data-example.csv.
  • Note that this dataset will not replicate results from the paper; it is intended only to illustrate the data structure.
• The scripts we use to clean, case control sample, and analyze the stillbirth data are located in at code/00_stillbirth-data-processing.

Data preparation and cleaning.

To use the scripts to run our analysis, the user must download the raw data files.

1. A stata script at /Data/000-stata-script-to-create-new-variables.txt processes the data, creating variables used in the analysis. The user must save the output of this script in a .csv file at /Data/BIRTH_DEATH_DATA_2004-2006.csv. The workflow then proceeds as follows.

2. Run the script /code/00_stillbirth-data-processing/00a-stillbirth-process-raw-data.R to clean the raw data and create variables for stillbirth analysis.

• This file takes the input raw dataset /Data/BIRTH_DEATH_DATA_2004-2006.csv.
  • This dataset must be saved by the user, by downloading the original publicly accessible data, and assembling in a .csv file.
• This file outputs the processed dataset /Data/stillbirth-clean.csv.

3. Run the script /code/00_stillbirth-data-processing/00b-stillbirth-create-case-control-samples.R to create the case-control sample of the stillbirth data.

• This file takes the input file /Data/stillbirth-clean.csv (created in the previous step).
• This file outputs one file, /data/perinatal-mortality/stillbirth-data-casecontrol-50-for-paper.csv. This file is used in the remainder of the analysis.

Code execution for analysis and simulation.

The code for both stillbirth analysis and the simulation can be run by the following command.

bash tsbart-analysis-coderun.sh

This script runs several R scripts, all located in the ./code subfolder. These scripts can also be executed individually.

• table-01.R
  • Generates Table 1 for cohort characteristics.
• table-02.R
  • Repeats second simulation on 100 datasets to calculate log-likelihood.
  • Creates Table 2.
• figure-02-and-figure06.R
  • Creates Figure 2 and Figure 6 (in Appendix A1).
  • Figure 2 illustrates results of the second simulation, comparing tsBART and BART.
  • Figure 6 is example of tuning expected number of crossings.
• table-03-a_generate-data.R
• table-03-b_one-example-run.R
  • These two scripts illustrate how data sets for the first simulation are generated and fit.
  • This code does not replicate Table 3; Table 3 is the result of repeating this process on 500 datasets for each weighting scenario to calculate MSE and point-wise coverage for each method and weighting scenario combination.
• figure-03.R
  • Illustrates hazard data generated for first simulation, comparing coverage across methods.
• figure-04-and-table-04.R
  • Fits models to up-sampled stillbirth data.
  • Creates Figure 4 and Table 4.
  • The file ending in "-single-iteration.R" illustrates the fit for a single dataset instead of all five runs.
• figure-05.R
  • Fits models to stillbirth data.
  • Creates Figure 5 to show estimated stillbirth risk curves for each method.

Helper functions.

There are several helper functions in the analysis that are used for cleaning data, assembling tables, etc. These are not included in the tsbart package, but are available in /code/helper-functions/. (These are sourced by other scripts as needed.)