This repository contains an R-based workflow for the ADS503 BMI Challenge. It ingests National Health Interview Survey (NHIS) data and linked mortality files from the CDC, preprocesses the data and fits a variety of machine learning models. The project uses the targets package with the crew backend to ensure reproducible pipelines.

BMI is not a great predictor of health outcomes and carries a lot of social stigma, so the goal of this project is to create a better model to understand ones risk for certain health outcomes using easily accessible data

• R (version 4.0 or later)
• Packages listed in _targets.R: targets, crew, tidyverse, arrow, SAScii, VIM, caret, pROC, earth

• All package dependencies and versions are recorded and managed with renv
• To restore the exact environment used for this analysis:

install.packages("renv")
renv::restore()

• See sessionInfo.txt for details of the analysis environment.

Clone the repository and execute the workflow from the project root:

git clone <repo-url>
cd ADS503-NHIS-BMI-Project
Rscript -e "targets::tar_make()"

Downloading the NHIS data directly from the CDC can take awhile, so the first run of the pipeline uses the raw_data in the repository, links it with the mortality data, then processes the data and trains models. All data is publicly available. Subsequent runs reuse the stored results. The specific code for downloading NHIS data directly from the CDC can be found in the ingest directory.

You can run individual stages by specifying target names, for example:

library(targets)
# only train BMI-only models
tar_make(names = "solo_bmi")

Results can be loaded back into R with tar_read():

library(targets)
model_df <- tar_read(modeling_df)

For a full tutorial of how the targets package works, visit this link targets

Visualize the workflow with:

library(targets)  
tar_visnetwork()

This command launches an interactive network graph showing all targets defined in _targets.R.

For a higher level view of the important parts

tar_glimpse()

For a table view with more details

tar_manifest()

The main stages declared in _targets.R (and implemented under R/target_parts/) are:

1. ingest_nhis_data – download and filter NHIS survey data.
2. process_mortality_data – download mortality files and tidy them.
3. link_mort_data – link survey records with mortality data.
4. secondary_pre_process – combine columns with differing names across years.
5. cont_var_plots – produce exploratory plots of numeric predictors.
6. missing_value_imputation – impute missing values using custom rules and hot-deck methods.
7. modeling_preparation – final cleaning and conversion of variables to factors.
8. solo_bmi – fit baseline models using BMI alone.
9. initial_evaluation – train and evaluate a set of baseline models.
10. hyperparameter_tuning – perform cross-validated tuning for several algorithms..

All scripts used in the pipeline are found in the R directory

• NHIS data is sourced from the CDC and linked to the mortality data
• Depending on year, there are anywhere from 500 to 700+ total variables in the publicly available datasets. Descriptions and notes of the ones we chose are found in the variables folder
• Survey descriptions, linked mortality data information, and variable codes and summaries from a few of the years can be found in the documents directory
• Full set of variables used for first round of modeling (and subsequently filtered) can be found in R/ingest_data/variables_to_use.R. These can be adjust to your liking

Landau, W. (2025). The {targets} R package user manual. Eli Lilly and Company. https://books.ropensci.org/targets/

National Center for Health Statistics. (2000–2018). National Health Interview Survey (NHIS): Documentation. Centers for Disease Control and Prevention. https://www.cdc.gov/nchs/nhis/documentation/index.html

National Center for Health Statistics. (2000-2018). Linked Mortality Files. Centers for Disease Control and Prevention. https://www.cdc.gov/nchs/data-linkage/mortality.htm