Author: Mengyu Liu
Date: 6/13/2023
Executive Summary:
The fatal drug overdose crisis is a pressing issue that continues to impact communities throughout the U.S. Individual risk factors, social determinants of health measures and environmental factors have been related to drug overdose deaths. However, integrating diverse information and generating individual predictions of drug overdose deaths on a unified scale remains challenging. The persistence of health disparities in overdose rates emphasizes the need for tailored and potentially personalized prevention strategies. This research aims to address the general research question of improving risk prediction and risk stratification for drug overdose deaths. We develop a machine-learning framework for improved prediction and identify a subgroup of people who were “contextually vulnerable”.The first part of the thesis focuses on examining the impact of the built environment, social determinants of health measures, and aggregated risk from the built environment at the neighborhood level on drug overdose death locations in Miami-Dade County, Florida. Risk Terrain Modeling (RTM) is utilized to assess the place features risk factors associated with drug overdose deaths spatially within Miami-Dade County ZCTAs. Logistic and zero-inflated regression models are used to investigate the effects of incident-specific social determinants of health (IS-SDH) measures and aggregated risk measures separately, as well as simultaneously, on drug overdose death locations each year. The identified high-risk areas and place features from RTM can inform the allocation of treatment and prevention resources. Combining an aggregated neighborhood risk measure with IS-SDH measures enables the identification of drug overdose death locations in certain years.In the second part of the thesis, individual risk factors and estimated risk from RTM are combined to build a spatial generalized linear mixed model (GLMM) for generating individual predictions of drug overdose deaths. A "shifting subjects" algorithm is proposed, and customized GLMMs are developed using model averaging to improve predictions. The concept of "contextual vulnerability" is introduced to identify individuals most affected by changes in the environment, forming a distinct "contextual vulnerability" group (CVG). The customized GLMM and averaging yield the largest gains in prediction accuracy for individuals in the CVG compared to conventional logistic regression models. Simulation studies are conducted to compare different models, and the contextual vulnerability analysis is applied to real-world data on drug overdose deaths in Seattle, WA, successfully identifying the CVG in specific years. This identification of contextual vulnerabilities can significantly impact prevention efforts and facilitate the development of potentially personalized prevention strategies.
