Adapting Spatiotemporal Gaussian Process Regression for Multinomial Data

Abstract

Estimating risk factor exposures over time is crucial to measuring and evaluating progress on behavioral, environmental, and occupational risks. Risk factor levels and trends are also a critical input in calculating the burden of diseases of health outcomes caused by exposure to health risks. Many health risk factors are multinomial in nature or in practice, and require more nuanced statistical treatment than traditional statistical methods provide. Spatiotemporal Gaussian Process Regression (ST-GPR) is a time-series model used primarily to estimate risk factor exposure within the Global Burden of Disease Project. We expanded the existing model to more accurately account for the unique requirements of ordinal and nominal multinomial time-series data. The expanded multinomial model was evaluated on two risk factor datasets of 1) occupational categories and 2) vaccination coverage, and compared based on out-of-sample fit and data coverage. We found that multinomial adaptations to the existing model led to slightly worse out-of-sample fit, yet major improvements in the validity of uncertainty and significant reductions in computational time.