A Bayesian Framework for Modeling Neighborhood-Level Social Determinants of Health

Abstract

This dissertation develops and applies Bayesian methods to model the multi-dimensional structure of social determinants of health (SDoH) at the neighborhood level. It progresses from applied probabilistic clustering of SDoH to identify profiles and their association with health outcomes, to methodological advances in mixture modeling for proportional data with feature selection, and finally to an integrative probabilistic framework that captures domain-specific signals that highlight neighborhoods of actionable priorities for resource allocation or targeted interventions. Together, these contributions provide a flexible and interpretable tools for understanding complex social environments and informing targeted public health interventions. Chapter 1 introduces the motivation for modeling SDoH as a multidimensional construct and provides an overview of the proposed methodological framework. Chapter 2 examines the relationship between neighborhood conditions and endometrial cancer care using a model-based clustering approach. A Bayesian multivariate Bernoulli mixture model is applied to American Community Survey data to identify distinct neighborhood SDoH profiles across Massachusetts. These profiles are then linked to patient-level data from the Massachusetts Cancer Registry using Bayesian logistic regression to assess associations with guideline-concordant care. The results demonstrate how multidimensional neighborhood profiles can serve as interpretable exposures, revealing disparities in care and highlighting opportunities for targeted interventions. Chapter 3 introduces SAMBA-Mix a salient multivariate Beta mixture model for clustering proportional SDoH data. The model leverages a multivariate beta likelihood to preserve the natural scale of bounded variables and incorporates feature saliency to identify determinants that drive cluster separation. This approach improves interpretability and robustness by allowing cluster formation to be guided by the most informative variables while reducing the influence of irrelevant features. Using data from the U.S. Agency for Healthcare Research and Quality (AHRQ) SDoH database, we demonstrate the utility of the model for deriving and characterizing neighborhood-level SDoH profiles in Massachusetts. Chapter 4 presents BMISS (Bayesian Multi-dimensional Integrative Social Structure), an unsupervised hierarchical framework for modeling neighborhood SDoH. Rather than collapsing information into a single index, BMISS produces domain-specific probability profiles that characterize domain-specific priority signals capturing patterns of co-occurring social determinants relevant for targeted resource allocation. By jointly modeling feature-level relationships and latent dependence structures, the framework provides a more refined and interpretable representation of neighborhood conditions, supporting more informed and context-specific decision-making. In an application to data from the AHRQ SDoH database, the model identifies distinct multidimensional profiles of priority signals across neighborhoods in Massachusetts and New York. Chapter 5 provides concluding remarks and avenues for future research.