Date of Award

Spring 2021

Document Type

Open Access Dissertation


Environmental Health Sciences

First Advisor

Dwayne E. Porter


Drug overdose is a leading cause of unintentional death in the United States and has contributed significantly to a decline in life expectancy from 2015 to 2018. Overdose deaths, especially from opioids, have also been recognized in recent years as a significant public health issue. To address this public health problem, this study sought to identify neighborhood-level (e.g., block group) factors associated with drug overdose and develop a spatial model using machine learning (ML) algorithms to predict the likelihood or risk of drug overdoses across South Carolina. This study included block group level socio-demographic factors and drug use variables which may influence the incidence of drug overdose. In particular, this study developed a new index of access to measure spatial access to treatment facilities and incorporated these variables to assess the relationship between drug overdose and accessibility to the treatment centers. We explored different ML algorithms (e.g., XGBoost, Random Forest) to identify optimum predictors in each category. The categories were combined into a final ensemble predictive model that addressed spatial dependency. An evaluation was conducted to validate that the final model generalized well across the different datasets and geographical areas. Results of the study identified strong neighborhood-level predictors of a drug overdose, pinpointing the most critical neighborhood-level factor(s) that place a community at risk and protect communities from developing such problems. These factors included proportion of households receiving food stamps, households with income less than $35,000, high opioid prescription rates, smoking accessories expenditures, and low accessibility to opioid treatment programs and hospitals. The generalized error of spatial models did not increase considerably in spatial cross-validation compared to the error estimated from normal cross-validation. Our model also outperformed the geographic weighted regression method. Our Results show that variables regarding socio-demographic factors, drug use variables, and protective resources can assist in spatial drug overdose prediction. Our finding highlights several specific pathways toward community-level intervention targeted to a vulnerable population facing potentially high burdens of drug abuse and overdose.