Approximate Bayesian Kernel Machine Regression via Random Fourier Features for Estimating Joint Health Effects of Multiple Exposures

Abstract: Environmental epidemiology has traditionally focused on examining health effects of single exposures, more recently with adjustment for co-occurring exposures. Advancements in exposure assessments and statistical tools have enabled a shift towards studying multiple exposures and their combined health impacts. Bayesian Kernel Machine Regression (BKMR) is a popular approach to flexibly estimate the joint and nonlinear effects of multiple exposures. However, BKMR faces computation challenges for large datasets, as inverting the kernel repeatedly in Markov chain Monte Carlo (MCMC) algorithms can be time-consuming and often infeasible in practice. To address this issue, we propose a faster version of BKMR using supervised random Fourier features to approximate the Gaussian process. We use periodic functions as basis functions and this approximation re-frames the kernel machine regression into a linear mixed-effect model that facilitates computationally efficient estimation and prediction. Bayesian inference was conducted using MCMC with Hamiltonian Monte Carlo algorithms. Analytic code for implementing Fast BKMR was developed for R. Simulation studies demonstrated that this approximation method yields results comparable to the original Gaussian process while reducing the computation time by 29 to 99%, depending on the number of basis functions and sample sizes. Our approach is also more robust to kernel misspecification in some scenarios. Finally, we applied this approach to analyze over 270,000 birth records, examining associations between multiple ambient air pollutants and birthweight in Georgia.