Author ORCID Identifier
Semester
Spring
Date of Graduation
2023
Document Type
Dissertation
Degree Type
PhD
College
School of Public Health
Department
Epidemiology
Committee Chair
Gordon S. Smith
Committee Co-Chair
Caroline P. Groth
Committee Member
Brian Hendricks
Committee Member
Erin L. Winstanley
Abstract
Introduction: Injury fatality rates in the United States (US) decreased throughout the majority of the 20th century, mostly due to declining rates of occupational and motor vehicle injuries. However, near the beginning of the 21st century, fatal injury rates in the US began to increase. This is principally due to the nation’s opioid epidemic, which has been characterized by different epidemic “waves”, each driven by overdoses associated with specific substances. Given the temporally dynamic nature of US injury trends, this study aimed to explore the application of time series analysis to injury data in the US. First, rates of non-fatal occupational injuries treated in US emergency departments were assessed to determine if non-fatal occupational injury rates mirror the historic decline of fatal occupational injuries in the 20th and 21st centuries. Next, we explored the temporal shift from prescription to illicit opioid overdose deaths in West Virginia (WV) to elucidate the transition between the opioid epidemic’s first and second waves in the state with the highest fatality rates in the nation. Finally, we compared the forecasting performance of three time series models when applied to national US opioid overdose data to explore what time series approaches best predict future rates of overdose.
Methods: Study one assessed temporal trends in non-fatal occupational injuries treated in US emergency departments (EDs) using data the National Electronic Injury Surveillance System – Occupational Supplement (NEISS-Work) dataset. Descriptive statistics were used to assess annual injury rate estimates and monthly seasonality. Autoregressive integrated moving average (ARIMA) modeling was used to quantify trends in ED-treated occupational injury rate estimates while controlling for serial data correlation. Analyses were conducted both overall and stratified by injury event type. Study two used data from the Drug Enforcement Agency’s (DEA) Automation of Reports and Consolidated Orders System (ARCOS) database (accessed via The Washington Post) to determine when shipments of oxycodone and hydrocodone tablets to WV began decreasing; tablet shipments were measured both as dosage units and morphine milligram equivalents (MMEs). To identify the exact point when tablet shipments began decreasing, we used locally estimated scatterplot smoothing (LOESS). The point when total tablet shipments began decreasing was used as an intervention point in an interrupted time series analysis (ITSA) of prescription and illicit opioid overdose death rates calculated using data from the WV Forensic Drug Database (FDD), which collects drug death data from the WV Office of the Chief Medical Examiner. Prescription opioid deaths were defined as those involving oxycodone or hydrocodone, while illicit opioid overdoses were defined as those involving heroin or synthetic opioids other than methadone. The ITSA impact of the LOESS-identified points was compared via Akaike Information Criteria (AIC) to that of the 2010 release of an abuse deterrent formulation (ADF) of OxyContin, which is widely cited as a driving factor initiating the transition between the opioid epidemic’s first and second waves. Study three examined the forecasting performance of ARIMA; Error, Trend, and Seasonality (ETS); and Facebook Prophet models when applied to national US opioid overdose death data, both overall and stratified by the type of opioid involved in overdoses. Overdose death counts were extracted from the Centers for Disease Control and Prevention Wide-ranging Online Data for Epidemiologic Research (CDC WONDER) database. Overdose death rates were calculated using monthly all-cause
ii
mortality as a denominator. Forecasts were validated used time series cross validation (TSCV), while forecast bias and predictive coverage probability were measured using mean average percent error (MAPE) and Winkler Scores, respectively.
Results: Study one found that US ED-treated non-fatal occupational injury rate estimates were highest in 2012 and lowest in 2019. Apart from falls, slips, and trips, all injuries occurred at the highest rate in a summer month. ARIMA modeling found that there was a significant decrease in monthly rate estimates for 2012-2019. Study two found that the point at which opioid tablet shipments (measured via dosage units) to WV began decreasing had a greater impact on changing rates of prescription and illicit opioid overdose rates than the 2010 ADF OxyContin release. Study three found that ETS models accurately forecasted monthly rates US opioid- involved overdoses while maintaining a high degree of precision relative to ARIMA or Facebook Prophet, particularly during the opioid epidemic’s fentanyl-dominated third wave.
Discussion: The findings presented here indicate that although occupational injury rates have likely continued their decades-long decline in the US, the nation’s opioid epidemic has contributed significantly to recent US injury rate increases and is temporally dynamic. Future research should explore trends in other injury data by expanding the methodology used here to other epidemiological contexts.
Recommended Citation
Lundstrom, Eric Wayne, "The Application of Time Series Analysis to Injury Epidemiology Data" (2023). Graduate Theses, Dissertations, and Problem Reports. 11869.
https://researchrepository.wvu.edu/etd/11869