Adaptive Gaussian Markov random field spatiotemporal models for infectious disease mapping and forecasting

Area of Science:

• Biostatistics
• Epidemiology
• Geospatial Statistics

Background:

• Existing literature uses adaptively parameterized spatiotemporal (ST) autoregressive (AR) or conditional autoregressive (CAR) models for Bayesian prediction of COVID-19 infection risks.
• These models aim to capture complex spatiotemporal dynamics and heterogeneities in infection risks.

Purpose of the Study:

• To synthesize, generalize, and unify ST AR and CAR model constructions augmented by adaptive Gaussian Markov random fields.
• To provide a flexible framework for disease mapping, spatial regression, and forecasting of areal-level disease risks.

Main Methods:

• Presentation of a general convolution construction for ST AR and CAR models.
• Development of illustrative models for characterizing local risk dependencies, modeling risk heterogeneities, and predicting/forecasting disease occurrences.
• Application to Bayesian hierarchical models for Poisson, zero-inflated Poisson, and Bernoulli data.

Main Results:

• A unified and generalized framework for spatiotemporal disease modeling is presented.
• The broadened constructions offer flexible parameterization options for disease mapping and spatial regression.
• The framework is applicable to quantifying covariate effects and forecasting infection occurrences and zero-infection probabilities.

Conclusions:

• The presented model constructions offer a flexible framework for modeling complex spatiotemporal data.
• The approach facilitates estimation, learning, and forecasting for disease surveillance and public health applications.
• This unified methodology enhances the prediction and forecasting of areal-level disease risks, including COVID-19.