This study was conducted using data from the Pennsylvania Trauma Outcome Study (PTOS) on patients admitted to Pennsylvania trauma centers between 2000 and 2009. The Pennsylvania population is representative of injured patients nationally (Mohan et al. 2011). The PTOS database is a population-based statewide trauma registry that includes data on all patients admitted with traumatic injuries to accredited trauma centers in Pennsylvania meeting one or more of the PTOS inclusion criteria: admission to the Intensive Care Unit or step-down unit, hospital length-of-stay greater than 48 hours, hospital admissions transferred from another hospital, and transfers out to an accredited trauma center (2007). The PTOS database includes de-identified data on patient demographics, Abbreviated Injury Score (AIS) codes and ICD-9-CM codes, mechanism of injury (based on ICD-9-CM Ecodes), comorbidities, physiology information, mechanisms of injury, in-hospital mortality and complications, transfer status, processes-of-care, and encrypted hospital identifiers. PTOS includes encrypted hospital identifiers, but it does not contain information on hospital characteristics and cannot be linked to other databases. Steps to insure data quality in the PTOS registry include the use of standard abstraction software with automatic data checks, a data definition manual, and internal and external data auditing (Rogers et al. 2011).
The analysis was limited to trauma patients with age greater than 16, excluding patients with burns, hypothermia, isolated hip fractures, superficial injuries, unspecified injuries, and nontraumatic mechanism of injury admitted to either Level I or Level II trauma centers. From this initial cohort of 226,283 patient observations, we excluded patients with missing information on transfer status (286), demographics (170); invalid AIS codes (12,662), race (15,515); and patients transferred out (4,012). We limited the analysis to black and white patients and excluded 1,751 Asian patients. (The race variable designated patients as either white, black, or Asian.) The study cohort consisted of 191,887 patients admitted to 28 Level 1 or Level II trauma centers. This study was approved by the institutional review board at the University of Rochester School of Medicine.
The main outcomes were (1) in-hospital death, (2) in-hospital death or major complication, and (3) failure-to-rescue. We defined this composite complication outcome if any of the following occurred after hospital admission: death, ARDS, acute myocardial infarction, acute respiratory failure requiring more than 48 hours of ventilatory support after a period of normal nonassisted breathing (minimum of 48 hours) or reintubation, aspiration pneumonia, pneumonia, pulmonary embolism, fat embolism syndrome, acute renal failure, central nervous system infection, progression of original neurologic insult, liver failure, sepsis, septicemia, empyema, dehiscence, gastrointestinal bleeding, small bowel obstruction, compartment syndrome, arterial occlusion, and postoperative hemorrhage. We used death or major complication as a composite outcome, as opposed to complication alone, because it is possible that low-quality care could result in high death rates and paradoxically low complication rates.
The unit of analysis was the patient. The independent variable was race (black vs. white). Trauma centers were stratified into quartiles based on the proportion of black trauma patients they treated: low (≤3.0 percent), medium (3.1–5.5 percent), medium-high (5.6–20.0 percent), and high (> 20 percent).
In our baseline model, we first estimated the independent effect of race on in-hospital death. We used the previously validated Trauma Mortality Prediction Model (TMPM-AIS) (Osler et al. 2008), modified by the addition of age, gender, comorbidities, mechanism of injury (based on Ecodes), transfer status, the GCS motor component, and systolic blood pressure, to control for confounding. Backward stepwise selection was used to select comorbidities. Fractional polynomial analysis was used to determine the optimal specification for continuous predictor variables (Royston and Altman 1994). To estimate the “within hospital” outcome disparity, we modified the baseline model by including hospital indicator variables as fixed effects. This fixed-effect model controls for patient and hospital-level confounders, so that the estimated race effect reflects the extent to which blacks and whites have different outcomes within the same hospital. We performed stratified analyzes after dividing patients into two groups based on injury severity (Baker et al. 1974): (1) mild-to-moderate injury severity (Injury Severity Score [ISS] <15); and (2) severe injury (ISS ≥15). To examine the effect of hospital choice on disparities, we re-estimated the baseline model which was modified to include the hospital proportion of black patients as a hospital-level factor. We then examined the interaction between hospital concentration of black patients and patient race by adding an interaction between hospital strata and race. We performed sensitivity analyzes in which we excluded patients transferred in from other hospital and patients who may have been dead-on-arrival (DOA). As there is no field to identify DOA patients in PTOS, we defined DOA as expired patients with short ED length of stays (<30 minutes) with a blood pressure of zero on admission to the emergency department. We also examined the impact of adding an indicator variable for trauma center status. The results of the sensitivity analyzes were similar to the primary analyzes and are not presented here.
Finally, to further clarify the association between hospital quality and the hospital proportion of black trauma patients, we re-estimated the modified TMPM-AIS using hospital random-effects. This model did not include race as a predictor variable. The empirical-Bayes estimate of the hospital effect was exponentiated to obtain the adjusted mortality odds ratio for each hospital (Glance et al. 2010). Hospitals with an adjusted odds ratio greater than 1 and whose 95 percent confidence interval did not include 1 were classified as low-quality outliers, whereas hospitals with adjusted odds ratios significantly less than 1 were classified as high-quality outliers. Caterpillar graphs were constructed to depict hospital quality as a function of hospital strata.
We then examined the time trend in the black–white difference in mortality by including year as a fixed effect in the baseline model, using the parameterization described by DeLong et al. (1997). To flexibly specify the year-race interaction, we used separate terms for the time variable for whites and for blacks. We tested the significance of the race-year interaction using a model in which year was specified as a continuous variable. Robust variance estimators were used due to the clustering of observations within hospitals (White 1980). The performance of TMPM-AIS was examined using measures of discrimination (C statistic) and calibration (calibration curves).
Analyzes were repeated using (1) the composite outcome of death or major complication, and (2) failure-to-rescue as the outcome of interest. The STATA implementation of the MICE method of multiple imputation described by van Buuren was used to impute missing values of the motor component of the GCS and the systolic blood pressure (van Buuren, Boshuizen, and Knook 1999). Model parameters estimated in the five imputed data sets were combined using Rubin's rule (Rubin 1987). Data management and statistical analyzes were performed using STATA SE/MP Version 11.1 (STATA Corp., College Station, TX, USA). All statistical tests were two-tailed and p values less than .05 were considered significant.