Integrating exploratory GAMs into the design of Bayesian BYM2 spatio-temporal models for small-country disease mapping

Abstract

Statistical disease mapping is central to public health surveillance, yet small-country settings with few regions and relatively short time series pose challenges for flexible yet identifiable spatio-temporal models. This article proposes a practical two-stage workflow that integrates generalized additive models (GAMs) with Bayesian spatio-temporal disease mapping, tailored to surveillance of acute respiratory infections, enteric infections and similar outcomes. In the first stage, we fit exploratory GAMs with smooth terms for calendar time, seasonality, spatial location and key covariates. Partial-effect plots and residual diagnostics are used to characterise non-linear relationships, seasonal patterns and spatial gradients, and to identify residual dependence that may require explicit modelling. In the second stage, we formalize these insights into a Bayesian latent Gaussian model fitted with INLA, combining a BYM2 spatial component, random-walk (RW1/RW2) temporal effects, optional cyclic seasonal terms and, where warranted, parsimonious spatio-temporal interactions. Penalised complexity (PC) priors regularise spatial and temporal hyperparameters, encouraging simple structure while allowing genuine variation to emerge. Lessons from small-country applications highlight the importance of balancing model complexity and identifiability, exploiting GAMs as a design tool rather than an endpoint, and using PC priors to stabilise inference when information is limited. The workflow offers a concrete template for combining exploratory flexibility and hierarchical regularisation in disease mapping for data-constrained surveillance systems, particularly in settings with few administrative regions and relatively short time series.