Semiparametric discovery and estimation of interaction in mixed exposures using stochastic interventions

Area of Science:

• Environmental Health Sciences
• Biostatistics
• Computational Biology

Background:

• Assessing combined health impacts of multiple environmental exposures is challenging.
• Existing methods struggle with complex interactions in high-dimensional data.
• Accurate analysis of exposure mixtures is crucial for public health.

Purpose of the Study:

• Introduce InterXshift, a novel semiparametric method for analyzing interactions in mixed environmental exposures.
• Facilitate discovery and efficient estimation of interaction effects using a nonparametric definition.
• Provide a robust tool for understanding complex exposure dynamics and their health outcomes.

Main Methods:

• Leverages stochastic shift interventions and ensemble machine learning.
• Employs a model-independent target parameter estimated via targeted maximum likelihood estimation (TMLE) and cross-validation.
• Contrasts outcomes from joint versus individual exposure interventions to detect synergistic/antagonistic effects.

Main Results:

• InterXshift effectively identifies true interaction directions and significant impacts, validated by simulations and NIEHS Mixtures Workshop data.
• Demonstrated efficacy in analyzing high-dimensional datasets for multi-exposure interactions.
• Applied to NHANES data to investigate furan exposure's interaction with leukocyte telomere length.

Conclusions:

• InterXshift offers significant methodological improvements for environmental health research.
• Enhances the analysis of complex exposure dynamics in large datasets.
• An open-source R package is available to facilitate community adoption and application.