|
Abstract:
|
Observational studies often collect information on multiple correlated exposures simultaneously (e.g., air pollutants), resulting in a multivariate exposure mixture with components that may impact disease. Linear models may not adequately capture the resulting mixture effect. Also, specifying a confounding set that satisfies causal identifiability assumptions with mixtures can be challenging, due to the numerous possible relationships between the exposures, confounders, and outcome. We propose an approach that combines the flexible modeling of kernel machine regression with a Bayesian approach to confounder adjustment. The kernel machine regression model for the exposures allows for non-linear and non-additive interactive effects, while Bayesian model averaging permits inference about causal effects that accounts for model uncertainty. Simulations demonstrate the method’s ability to produce unbiased estimates for the mixture effect in a variety of scenarios (e.g., high and low correlation among exposures; linear and non-linear outcome models). We apply the method to data from the Sister Study to estimate the effect of metal concentrations in the body on BMI in a cohort of women.
|
