Data sources
Data were obtained from NEDS for 2006–2012. Based on a 20% stratified single cluster sample of hospital-based emergency departments (EDs), NEDS is the largest most representative single publicly available ED database in the US. Core files consist of 100% of annual visits from sampled community hospitals, defined as non-federal, general, short-term, and specialty hospitals (such as pediatric hospitals and academic medical centers). NEDS utilizes a stratified sampling strategy based on geographic area, urban/rural area, ownership (i.e., government, private, not-for-profit), trauma center and teaching status, and bed size (Agency for Healthcare Research and Quality Healthcare Cost and Utilization Project (HCUP) 2016). For population-based rates, annual national and regional population estimates were obtained from HCUP, originating from the US Census Bureau. For hospital discharge-based rates, we used the universe of weighted NEDS ED visits. For travel-based rates, estimated national vehicle miles traveled were obtained from annual highway statistics maintained by the Federal Highway Administration (US Department of Transportation, Federal Highway Administration 2017).
Inclusion criteria and study measures
We used R and MonetDB to read in the full initial dataset of 198,102,435 unweighted observations. Age groups were defined to be clinically relevant and consistent with available population estimates, as follows: 0–4, 5–9, 10–14, and 15–19 years. Four International Classification of Diseases, Ninth revision, Clinical Modification (ICD-9-CM) coded external cause of injury (E code) variables were used to identify injuries to pedestrians or pedalcyclists being struck by or involved in a collision with a motorized vehicle. We use the common term “bicyclist” throughout the remainder of this study to refer to all types of non-motorized pedalcyclists. We excluded E codes for bicyclists injuring other bicyclists or themselves (E8261–4, E8268–9), and injuries to drivers and motor vehicle occupants involved in a collision with a pedestrian (E8140–44). Pedestrians struck by bicyclists were included, given the potential resemblance in injury mechanism with other pedestrian-vehicle crashes. A detailed list of codes is available as an Additional file 1.
Injury severity, or probability of survival, was quantified using the ICD-derived Injury Severity Score (ICISS) as proposed by Osler et al. (Osler et al. 1996). First, survival risk ratios (SRRs) for each injury diagnosis were “...calculated as the ratio of the number of times a given ICD-9-CM code occurs in (surviving patients) to the total number of occurrences of that code.” Second, “the product of all the survival risk ratios (was computed) for each of an individual patient’s injuries” for up to ten different injuries (Segui-Gomez and Lopez-Valdes 2012). ICISS was then defined as the probability of patient survival and ranges from 0 to 1. As ICISS is, perhaps, most useful as a dichotomous indicator, (Stevenson et al. 2001) we defined severe injury as a survival probability <94% (i.e., a 6% or greater probability of death), as proposed by Gedeborg (Gedeborg 2009). In our previous analyses of trauma mortality, this cut point (0.94) produced a odds ratio of 6.75 (95% CI 6.48, 7.03) in multivariable logistic regression (DiMaggio et al. 2016).
Fatal outcomes were defined as patient records having a disposition of death in the ED or in the hospital, if admitted. Trauma center designations were based on the AHRQ “HOSP TRAUMA” indicator variable found in NEDS, which comes from the Trauma Information Exchange Program database (TIEP). Primary ICD-9-CM codes were categorized by the Barell Matrix, (Fingerhut et al. 2002; Barell et al. 2002) an injury diagnosis tool used internationally to standardize the classification of ICD-9-CM injury codes according to 12 nature-of-injury columns and 36 body-location rows. We also explored TBI severity using the 2002 Barell Matrix definition, as Type I (intercranial injury with moderate or prolonged loss of consciousness), Type II (no intercranial injury, loss of consciousness less than one hour), or Type III (no intercranial injury and no loss of consciousness).
Statistical analysis
Descriptive epidemiology, including visit counts and population-based rates, age, gender, trauma center status, injury severity, and presence of TBI were estimated using survey-adjusted counts and means with the R package “sqlsurvey.” Ratio estimates and differences were calculated using the simulation method (Greenland 2004) based on survey-adjusted counts and standard errors, with each simulation consisting of 1000 random normal draws. Trends over time were evaluated with linear regression using year as the predictor variable.
Utilizing a retrospective cohort approach, we modeled the association of pedestrian vs. bicyclist injury with in-hospital fatality risk, controlling for age, gender, injury severity, trauma center status, and presence of TBI, with an interaction term for pedestrian status and TBI:
$$ \mathrm{death}={\upbeta}_{\mathrm{intercept}}+{\upbeta}_{\mathrm{age}}+{\upbeta}_{\mathrm{gender}}+{\upbeta}_{\mathrm{severity}}+{\upbeta}_{\mathrm{traumaCenter}}+{\upbeta}_{\mathrm{pedestrian}}+{\upbeta}_{\mathrm{TBI}}+{\upbeta}_{\mathrm{pedestrian}\cdotp \mathrm{TBI}} $$
Where, “age” was measured continuously in years, “gender” was an indicator variable for male (0) vs. female, “severity” was dichotomously coded 1 for ICISS <0.94 vs. ≥94 (0), “traumaCenter” was dichotomously coded 1 for treatment at a level 1 or 2 trauma center vs. non-trauma center (0), “pedestrian” was an indicator variable for pedestrian versus bicyclist (0), and TBI was dichotomously coded 1 for patients having traumatic brain injury as the primary diagnostic code per the Barell matrix, otherwise (0). The interaction term, “pedestrian.TBI” indicates the 4-level combination of the dichotomous variables for pedestrian status and TBI. Model and variable selection were based on descriptive results and univariable regression for the outcome of fatality. Modeling was conducted using the R “rms” package with robust covariances and options set to account for survey weighting and clustering. We tested for the assumption of linearity of the year variable and controlled for year-to-year variability in the survey results using an approach recommended by the Centers for Disease Control and Prevention (Centers for Disease Control and Prevention 2017; Bieler et al. 2010).
We assessed additive interaction of pedestrian status and TBI on fatality risk using a component-cause approach described by Darroch (Darroch 1997) and Rothman and Greenland (Rothman and Greenland 1998). To determine if the observed risk of pedestrian injury and TBI (Rped + TBI+) exceeds what we might expect if the two risks did not interact, we set up an interaction contrast (IC) as follows:
$$ \mathrm{IC}=\left({\mathrm{Risk}}_{\mathrm{ped}+\mathrm{TBI}+}-{\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}-}\right)-\left({\mathrm{Risk}}_{\mathrm{ped}+\mathrm{TBI}-}-{\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}-}\right)-\left({\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}+}\kern0.5em -\kern0.5em {\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}-}\right)={\mathrm{Risk}}_{\mathrm{ped}+\mathrm{TBI}+}\kern0.5em -{\mathrm{Risk}}_{\mathrm{ped}+\mathrm{TBI}-}\kern0.5em -{\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}+}\kern0.5em +{\mathrm{Risk}}_{\mathrm{ped}-\mathrm{TBI}-} $$
Interaction is considered present if IC ≠ 0, i.e., the combined effect does not equal the sum of the individual pedestrian and TBI effects. The proportion of excess risk attributable to the interaction between pedestrian injury and TBI was calculated as:
$$ \mathrm{IC}/{\mathrm{Risk}}_{\mathrm{ped}+\mathrm{TBI}+}. $$
We took into account survey variability by running bootstrap estimates of this equation and report the result as a point estimate with 95% confidence interval.
The study was approved by the New York University School of Medicine Institutional Review Board, and conforms to the STROBE statement for observational studies, excluding elements not applicable to a retrospective study design using administrative data. Detailed R code to reproduce or adapt our methods is available in Additional file 2.
