| dc.description.abstract | Statistical modeling has evolved from the classical Linear Model (LM) to more flexible frameworks, including the Generalized Linear Model (GLM), Linear Mixed Model (LMM), and Generalized Linear Mixed Model (GLMM), enabling the analysis of non-Gaussian, hierarchical, and spatio-temporal data. These methodological advances have become increasingly important in epidemiology, where infectious diseases such as dengue exhibit substantial spatial and temporal heterogeneity. Recent developments in remote sensing provide continuous environmental information for disease surveillance; however, satellite-derived environmental indices are indirect measurements affected by sensor limitations, atmospheric disturbances, cloud contamination, and image-processing procedures. Ignoring this uncertainty may bias parameter estimation and reduce predictive performance. Although measurement-error methods have been extensively investigated in classical statistical models, their integration into Bayesian spatio-temporal disease-mapping remains relatively limited. Therefore, this dissertation develops a unified Bayesian spatio-temporal Poisson framework that explicitly accounts for measurement error in remote-sensing covariates for dengue modeling.
The proposed framework extends the Bayesian hierarchical GLMM by jointly modeling fixed effects, localized spatial dependence, temporal dependence, spatio-temporal interaction, and covariate measurement error within a single coherent inferential framework. Spatial autocorrelation is represented by Localized Conditional Autoregressive (LCAR) priors, while temporal dependence and spatio-temporal interactions are incorporated through hierarchical random effects. This hierarchical Bayesian formulation enables simultaneous estimation of all model components while propagating uncertainty across multiple levels of the model, providing a flexible and statistically rigorous approach for analyzing complex spatio-temporal epidemiological data.
The first study demonstrated that Sentinel-2-derived environmental indices provide reliable environmental proxies for explaining dengue incidence. The study compared four statistical models, including conventional Poisson regression, clustered Poisson regression, localized spatial Poisson regression, and Hierarchical Bayesian Spatio-Temporal Localized Conditional Autoregressive (HBSTLCAR) model. The results showed that explicitly accounting for spatial dependence substantially improved predictive performance and more effectively captured spatial heterogeneity than conventional regression models.
The second study incorporated measurement-error correction into a clustered Poisson regression model and demonstrated that explicitly accounting for uncertainty in remotely sensed environmental covariates improved both parameter estimation and predictive accuracy. These findings indicate that satellite-derived environmental variables should not be regarded as error-free measurements when used in epidemiological modeling.
The third study developed the proposed HBSTLCAR Model with Measurement Error (HBSTLCAR–ME) and evaluated its performance through both simulation and empirical studies. The simulation considered nine scenarios generated from combinations of three levels of spatial dependence (none, moderate, and high) with three levels of temporal dependence (none, moderate, and high). Across nearly all scenarios, HBSTLCAR-ME consistently outperformed the corresponding model without measurement error correction, yielding lower prediction errors and improved Bayesian model-fit criteria. The simulation further demonstrated that the effectiveness of measurement-error correction depended on the underlying spatio-temporal structure, with greatest improvements observed under moderate temporal dependence. Under strong spatial dependence, the performance gap between the two models become smaller because localized spatial random effects explained much of the residual variation. In the empirical application, the proposed HBSTLCAR-ME model achieved the best overall predictive performance while preserving localized spatial and temporal risk structures. Measurement-error correction primarily improved estimation of environmental covariate effects without materially altering the underlying spatial dependence, indicating that localized spatial heterogeneity remained the primary source of variation, whereas measurement error constituted an important secondary source of uncertainty.
The principal contribution of this dissertation is the development and comprehensive validation of a unified Bayesian spatio-temporal Poisson framework that integrates localized spatial dependence, temporal correlation, spatio-temporal interaction, and measurement error in remotely sensed environmental covariates. Validation through both simulation studies and empirical epidemiological applications demonstrates that explicitly accounting for measurement uncertainty consistently improves predictive performance, enhances parameter estimation, and strengthens Bayesian inference while preserving the underlying spatio-temporal dependence structure. Furthermore, the proposed framework establishes that remotely sensed environmental indices can serve as reliable proxy variables for dengue surveillance, provided that their inherent measurement uncertainty is explicitly accommodated within a Bayesian spatio-temporal modeling framework. | |
