Title Information
Type of Resource (primo): dissertations
Name: Personal
Name: Personal
Name: Personal
Name: Personal
Name: Personal
Name: Corporate
Origin Information
Physical Description
Note: thesis: Thesis (Ph. D.)--Brown University, 2026
Genre (aat): theses
Abstract: Consistent medical care is essential for the health of people living with HIV (PLWH). Regular engagement in care improves access to antiretroviral therapy (ART), reduces progression to Acquired Immune Deficiency Syndrome (AIDS), and is associated with improved survival rates compared to PLWH who do not receive regular medical care. Population-level engagement in HIV care is commonly summarized through the HIV care cascade, a conceptual framework that defines key benchmarks for monitoring the effectiveness of HIV healthcare systems and a guideline for identifying gaps in care. The objective of this dissertation is to develop and implement data-driven tools to predict retention in HIV care to support clinical decision making at the Academic Model Providing Access to Healthcare (AMPATH), a large network of clinics providing HIV care in western Kenya. Currently, AMPATH dedicates considerable resources to patient outreach following a missed visit; however, the human resources used for outreach are not unlimited. Advanced outreach conducted prior to a scheduled return visit may increase the likelihood of continued engagement in care, motivating the development of models that predict a patient’s risk of missing an upcoming scheduled clinic visit. The first aim develops a principled statistical framework for predicting visit-level disengagement that explicitly aligns both model formulation and sampling strategies with the underlying data-generating process at each scheduled return visit. This framework uses a flexible Bayesian multinomial model that accounts for competing risks in the outcome space while simultaneously addressing missingness in the covariate space. The second aim focuses on interpretability by developing methods for local variable importance to support personalized outreach. We introduce a sampling-based approach grounded in predictive mean matching that samples from appropriate conditional distributions. We conduct simulations to demonstrate that this approach is scalable, robust to model misspecification, and applicable across a wide range of covariate types. The third aim examines how missing data at both model development and deployment affects global variable importance estimation. We extend a model-agnostic approach, Leave Out Covariates (LOCO), to explicitly account for missingness and use extensive simulations to evaluate its properties.
Subject
Subject
Subject
Subject
Subject
Language
Record Information