BARTSIMP: Flexible spatial covariate modeling and prediction using Bayesian Additive Regression Trees

Area of Science:

• Spatial Statistics
• Machine Learning
• Biostatistics

Background:

• Accurate prediction in spatial statistics is crucial.
• Incorporating spatial covariates enhances predictive performance.
• Existing methods lack reliable uncertainty estimates for spatial data.

Purpose of the Study:

• To develop flexible regression models for spatial data with nonlinearities and interactions.
• To combine Gaussian process models with Bayesian Additive Regression Trees (BART).
• To address limitations in uncertainty estimation from current machine learning approaches.

Main Methods:

• Integration of Gaussian process spatial models with Bayesian Additive Regression Trees (BART).
• Utilizing Markov chain Monte Carlo (MCMC) and Integrated Nested Laplace Approximation (INLA) for computational efficiency.
• Simulation studies to evaluate method performance.
• Application to anthropometric response prediction using complex survey data from Kenya.

Main Results:

• The proposed model demonstrates improved predictive performance.
• Reliable uncertainty estimates are provided, overcoming limitations of existing methods.
• Successful application to complex survey data, accounting for sampling design.

Conclusions:

• The combined Gaussian process and BART model offers a powerful tool for spatial prediction.
• The method effectively handles nonlinearities, interactions, and spatial dependence.
• This approach provides accurate predictions with reliable uncertainty quantification for complex spatial datasets.