IDENTIFYING MAIN EFFECTS AND INTERACTIONS AMONG EXPOSURES USING GAUSSIAN PROCESSES.

Area of Science:

• Environmental Health
• Biostatistics
• Statistical Modeling

Background:

• Assessing the joint impact of chemical exposures on human health is complex.
• Existing methods often lack interpretability or struggle with model selection for main effects and interactions.
• Semiparametric models are needed to disentangle linear, interaction, and nonlinear health outcome components.

Purpose of the Study:

• To develop a flexible Bayesian framework (MixSelect) for selecting important main effects and interactions of chemical exposures on health outcomes.
• To address challenges in model interpretability, uncertainty, and nonidentifiability in semiparametric models.
• To provide a method for understanding the complex relationships between environmental factors and human health.

Main Methods:

• A Bayesian approach using variable selection priors and a Markov chain Monte Carlo (MCMC) algorithm.
• Incorporation of a heredity constraint to ensure interactions are only considered with main effects.
• Adaptation of a projection approach for Gaussian process modeling and dimension reduction for efficient sampling.

Main Results:

• The MixSelect framework effectively performs variable selection for main effects, pairwise interactions, and nonlinear deviations in health outcome models.
• The proposed methods successfully address nonidentifiability issues between linear and nonparametric components.
• Simulation studies and analysis of National Health and Nutrition Examination Survey (NHANES) data demonstrate the framework's utility.

Conclusions:

• The MixSelect framework offers a robust and interpretable approach to analyzing the joint effects of chemical exposures on health.
• This method enhances the understanding of complex environmental health relationships by identifying significant effects and interactions.
• The developed Bayesian methodology provides a valuable tool for public health research and policy.