A Bayesian Hierarchical Framework for Modeling Multi-Source Public Health Data

Thesis Abstract

The increasing availability of diverse public health data sources, including electronic health records, survey data, national health registries, and environmental monitoring systems, has created opportunities for comprehensive health surveillance but has also introduced complex challenges in data integration and analysis. These challenges stem from heterogeneity in data structure, varying levels of data quality, and the need for sophisticated statistical models capable of capturing dependencies across multiple sources. This study aims to develop and evaluate a Bayesian hierarchical framework tailored for modeling multi-source public health data, with the primary objective of improving the accuracy and interpretability of health outcome estimations across different geographical and temporal scales. The specific objectives include formulating a flexible Bayesian hierarchical model that accounts for source-specific biases and uncertainties, implementing computational algorithms such as Markov Chain Monte Carlo (MCMC) for parameter estimation, and assessing the model’s performance using simulated and real-world datasets. The research adopts a quantitative, methodological design grounded in Bayesian statistical principles and leverages a multilevel modeling approach to synthesize data from diverse sources. The target population comprises aggregated public health datasets collected across multiple regions, including healthcare utilization data for a chronic disease prevalence study involving approximately 50,000 individual records from hospital and clinic sources over a five-year period. Data collection instruments encompass structured data extraction protocols from electronic health systems, standardized survey questionnaires, and publicly available environmental monitoring reports. To ensure the validity and reliability of the data, multi-stage validation techniques, including consistency checks, calibration procedures, and sensitivity analyses, are employed throughout the data handling process. Analytically, the study adopts Bayesian hierarchical modeling techniques, integrating prior distributions informed by epidemiological theories such as the Social Determinants of Health framework and the Ecological Model. Model specification involves hierarchical levels accounting for individual, community, and regional effects, with source-specific bias parameters incorporated as random effects. MCMC algorithms are utilized for posterior inference, with convergence diagnostics such as Gelman-Rubin statistics and trace plots ensuring the robustness of parameter estimates. Model comparison metrics, including Deviance Information Criterion (DIC) and Watanabe-Akaike Information Criterion (WAIC), guide the selection of optimal models. Additional analytical tools include residual analysis and posterior predictive checks to evaluate model fit and predictive performance. The anticipated findings include a robust Bayesian hierarchical model capable of effectively integrating heterogeneous data sources, with improved accuracy in estimating disease prevalence and risk factors at multiple levels of aggregation. The model is expected to demonstrate superior performance over traditional single-source methods, particularly in adjusting for source biases and handling missing data through the hierarchical structure. The results will offer critical insights into spatial and temporal patterns of health outcomes, emphasizing the importance of multi-source data fusion in public health surveillance and policymaking. This study significantly advances statistical methodology in public health research by providing a comprehensive Bayesian framework adaptable to various health data integration challenges. It contributes to the existing literature by operationalizing hierarchical models that explicitly account for source heterogeneity and bias, thereby enhancing the credibility and utility of combined datasets in epidemiological investigations. The practical implications include informing policymakers on more reliable health indicators and fostering more precise resource allocation based on integrated data analyses. In conclusion, the study recommends adopting Bayesian hierarchical models as standard analytical tools for multi-source public health data. It underscores the need for continued methodological development to incorporate real-time data streams and machine learning techniques, and emphasizes capacity-building for health practitioners in Bayesian analytics to support evidence-based decision-making. Future research directions include extending the framework to incorporate spatial-temporal dynamics and developing user-friendly software packages to facilitate broader application across various public health domains.