This article presents a comprehensive and methodologically rigorous investigation into the spatio-temporal dynamics of dengue transmission in Brazil, framed within an advanced quantitative epidemiological paradigm. It proposes and validates an integrated, multi-layer analytical architecture that systematically reconciles causal inference, spatial heterogeneity, temporal dependence, and high-performance prediction at the national scale. Drawing upon a decade-long municipal panel dataset encompassing all 5,570 Brazilian municipalities (2010–2019), the study operationalizes dengue incidence as a smoothed epidemiological outcome and explicitly models its association with a broad set of socioeconomic, demographic, housing, sanitation, and structural determinants derived from authoritative national databases. The analytical strategy is deliberately hierarchical and epistemologically plural, progressing from robust exploratory spatial data analysis to global parametric modeling, local non-stationary regression, Bayesian spatio-temporal hierarchical inference, state-of-the-art machine learning, and ensemble learning. At its inferential core, the study employs Bayesian spatio-temporal models estimated via Integrated Nested Laplace Approximation (INLA), incorporating structured spatial random effects and temporally correlated processes to rigorously control for latent spatial dependence and temporal autocorrelation. This framework enables the estimation of interpretable relative risks with full uncertainty quantification, thereby providing statistically defensible insights into the structural determinants of dengue transmission. Complementarily, Geographically Weighted Regression (GWR) is used to formally test and map spatial non-stationarity, revealing the inherently local and context-dependent nature of socioeconomic effects across the Brazilian territory. To address the distinct demands of prediction and operational surveillance, the study integrates non-parametric machine learning through extreme gradient boosting (XGBoost), optimized via Bayesian hyperparameter search, and further enhances predictive accuracy through a stacking-based ensemble that synthesizes information from Bayesian, local, and machine-learning models. Model validation is conducted under stringent spatial and temporal cross-validation schemes, ensuring genuine out-of-sample generalizability. The results demonstrate that dengue incidence in Brazil is governed by strongly non-stationary socioeconomic processes embedded in spatially structured contexts, and that models neglecting these properties are fundamentally misspecified. The proposed integrated framework achieves unprecedented predictive performance for a nationwide public health phenomenon while retaining inferential robustness and interpretability. The study culminates in high-resolution predictive risk maps with quantified uncertainty, explicitly designed to support precision health surveillance, territorialized intervention planning, and evidence-based allocation of public resources. By unifying Bayesian inference, spatial statistics, and machine learning within a single coherent methodological architecture, this work establishes a new benchmark for quantitative modeling of neglected infectious diseases and advances the epistemological foundations of modern health surveillance.
Caio A. Rocha (Sat,) studied this question.