We used data from the North Carolina Trauma Registry (NCTR), a statewide registry and cooperative effort between eighteen North Carolina hospitals, including all 17 trauma centers in the state (6 Level I, 3 Level II, and 8 Level III hospitals), and the North Carolina Office of Emergency Medical Services (Thomason 2008; Office of Emergency Medical Services). The NCTR has been in place since 1987 and collects near real-time information using standardized data definitions based off of the National Trauma Registry of the American College of Surgeons and designated chart abstractors. All hospitalizations where a patient is diagnosed with a traumatic injury (ICD-10-CM: S00-S99, T07, T14, T20-T28, T30-T32, T71, T79.A1-T79.A9), and is admitted to the hospital, taken to the operating room from the emergency department, transferred, or dies due to their injury are included. Unplanned readmissions within 30 days of the initial injury are also included.
For this study, we included all assault hospitalizations that occurred between January 1, 2019, and December 31, 2020. Both GSW and non-GSW assault hospitalizations were identified using the ICD-10-CM code framework from the National Center for Health Statistics and National Center for Injury Prevention and Control (Hedegaard et al. 2019). North Carolina population counts for 2019 were used to estimate the weekly hospitalization rates per 1,000,000 North Carolina residents. To account for variation between weekday and weekend admissions, we calculated the weekly rates of GSW and non-GSW assault hospitalizations between January 6, 2019, and December 26, 2020. Hospitalizations that occurred in partial weeks (i.e., January 1–5, 2019, and December 27–31, 2020) were excluded from our models to avoid introducing bias due to underestimating the total hospitalization rate for those weeks.
All patient demographics were abstracted from electronic medical records associated with the assault hospitalization by NCTR abstractors. Race and ethnicity were self-reported by the patient (or family member) if they were present and capable; otherwise, it was based on staff designation. Patients were categorized as non-Hispanic Black/African American, Hispanic/Latino, non-Hispanic White, and non-Hispanic other race; other race included individuals identified as American Indian, Asian, Pacific Islander, multiracial, or “other” race in their hospitalization record.
Differences in patient demographics and clinical characteristics among patients admitted for traumatic injuries between 2019 and 2020 were compared using standardized differences. An absolute difference > 0.20 was considered meaningful.
We then conducted a natural experiment using an interrupted time series design and segmented linear regression (Wagner et al. 2002; Taljaard et al. 2014). Interrupted times series is a powerful, quasi-experimental approach to evaluate longitudinal effects of interventions and policy changes that occur at a specific point in time. This approach is preferable to a more traditional “pre/post” test as it allows researchers to visualize both immediate (intercept) and gradual (slope) changes before and after the intervention, as opposed to just comparing the average rates before and after implementation (Wagner et al. 2002; Taljaard et al. 2014). These models can also incorporate multiple interventions within a single analysis, which is critical when assessing COVID-related policies. Using ordinary least squares, we ran race/ethnicity and sex/age-specific segmented linear regression models to estimate the trend in trauma hospitalization rates between each pair of COVID-related executive orders and policies.
The policies of interest included: U.S. declaration of a public health emergency (1/31/2020), statewide Stay-at-Home order (3/30/2020), partial lifting of Stay-at-Home restrictions (Phase 2: Safer-at-Home, 5/22/2020), and further lifting of restrictions (Phase 2.5: Safer-at-Home, 9/4/2020). Several statewide orders were not included in analyses because either the order made relatively small changes to existing orders (e.g., Phase 1 lifting of Stay-at-Home orders) or it occurred within several weeks of a prior order that we believed would be more salient (e.g., North Carolina declaring a state of emergency). A full list of executive orders and their effective dates are included in Additional file 1: Table S1.
To reduce error in our model, we used a transformed cosine periodic function to control for potential seasonal fluctuations in hospitalization rates.(Brookhart and Rothman 2008) To account for autocorrelation over time, we used Durbin–Watson tests (α = 0.05) to specify autoregressive parameters in our models for lags up to 60 weeks. Our models did not include parameters for level changes (i.e., intercept changes) to focus our analysis on an a priori-hypothesis that only gradual changes in injury hospitalization rates would be observed. Because no significant trend changes in rates were seen prior to the pandemic, the average weekly rate of hospitalizations for this time period was estimated by taking the mean of all estimated weekly rates prior to the U.S. declaration of a public health emergency. Post-policy slopes were calculated by summing the slope after the policy of interest (e.g., Stay-at-Home order) with the preceding slopes (e.g., pre-pandemic slope and slope after the U.S. declared a public health emergency) (Wagner et al. 2002; Taljaard et al. 2014). 95% confidence intervals were calculated using the approximate standard errors generated for each post-policy slope. We have used these methods previously (Strassle et al. 2022). Due to low rates, stratified models among women (all ages) and older men (≥ 65 years old) could not be performed.
All analyses were performed using SAS version 9.4 (SAS Inc., Cary, North Carolina). This study was deemed exempt by two Institutional Review Boards.
