A Tutorial on Conducting Mediation Analysis with Exposure Mixtures

This repository provides code and functions for conducting mediation analysis in the context of exposure mixtures, focusing on both simulated data and real-world applications from the PROTECT cohort study. The methods implemented include single exposure mediation analysis (SE-MA), principal component-based mediation analysis (PC-MA), environmental risk score-based mediation analysis (ERS-MA), and Bayesian kernel machine regression causal mediation analysis (BKMR-CMA).

The corresponding publication is available here. For a quick illustration of the SE-MA, PC-MA, ERS-MA, and BKMR-CMA methods applied to an example dataset, see Example.md.

Overview

In environmental health research, individuals are often exposed to complex mixtures of chemicals rather than isolated exposures. This project illustrates practical strategies for mediation analysis under such settings, addressing challenges such as multicollinearity, sparsity, and nonlinearity among exposures.

The codebase supports:

• Simulation of exposure mixtures and outcomes
• Application of mediation methods on simulated datasets
• Application of mediation methods on real-world PROTECT cohort data
• Comparison of different mediation analysis strategies

Data Availability

Due to the size of the datasets, they are hosted separately.
You can download the datasets here:
🔗 Download Data from Google Drive

After downloading, place the data files in the appropriate directories as indicated in the scripts.

Repository Structure

• Functions/
  • Functions_DataGen.R: Functions for simulation data generation
  • Functions_ERS.R: Functions for ERS-MA
  • Functions_IndTesting.R: Functions for single exposure mediation analysis (SE-MA)
  • Functions_PCA.R: Functions for PC-MA
• Evaluation/
  • data_sim.R: Master script for simulating 1 dataset with 100,000 observation under each degree of effect
  • estimate_true_ers.R: Estimates "true" ERS-MA effects using a large simulated dataset (n = 100,000)
  • estimate_true_pca.R: Estimates "true" PC-MA effects using a large simulated dataset (n = 100,000)
  • eva_bkmr.R: Computes variable selection accuracy for BKMR-CMA
  • eva_ers.R: Computes relative bias for ERS-MA
  • eva_pca.R: Computes relative bias for PC-MA
  • eva_sema.R: Computes relative bias and variable selection accuracy for SE-MA
  • functions_ERS.R: Functions for ERS-MA
  • plot_relative_bias.R: Generates comparative plots of percent relative bias across methods
  • plot_selection_metrics.R: Generates comparative plots of true/false positive rates across methods
  • summarize_sim_results.R: Extracts and stores relevant quantities from each simulated dataset
• BKMR_Protect.R: BKMR-CMA analysis on PROTECT dataset
• BKMR_sim.R: BKMR-CMA analysis on simulated dataset
• Data_Clean_PROTECT.R: Data cleaning and preprocessing for PROTECT dataset
• ERS_Protect.R: ERS-MAs on PROTECT dataset
• ERS_sim_Case1.R: ERS-MA on simulated data (case 1)
• ERS_sim_Case2.R: ERS-MA on simulated data (case 2)
• SEMA_Protect.R: Single exposure mediation analysis (SE-MA) on PROTECT dataset
• SEMA_sim.R: Single exposure mediation analysis (SE-MA) on simulated dataset
• PCA_Protect.R: PC-MA on PROTECT dataset
• PCA_sim.R: PC-MA on simulated dataset
• Simulation_Data_Generation.R: Script for generating simulated datasets
• Example.md: Applies SE-MA, PC-MA, ERS-MA, and BKMR-CMA to an example dataset
• Example_files: Output figure from Example.md
• LICENSE: License information
• README.md: Project documentation (this file)

Notes

• Supporting functions are located in the Functions/ directory and sourced within scripts.
• For computationally intensive methods such as BKMR, runtime may be substantial depending on your computing environment.

License

This project is licensed under the terms specified in the LICENSE file.