Study design
We conducted a retrospective analysis using statewide linked hospital, prehospital EMS, and mortality data for NJ. The study protocol was approved by the Institutional Review Boards (IRBs) at Rutgers University and the New Jersey Department of Health (DOH). Because the study was based on secondary data and the study team had no access to personal identifiers, written informed consent was not required by the IRBs.
Setting
New Jersey is a densely populated state of 8.8 million residents [12]. There are 73 acute care hospitals operating in the state, all of which are required by state law to maintain a full-service emergency department (ED) 24 h per day. The number of hospitals adopting TH grew from none in 2004 to 38 in 2010. The state’s prehospital emergency medical services (EMS) include a mix of career and volunteer basic life support (BLS) ambulance companies. In most communities, private or municipal BLS units are supplemented by 21 hospital-based advanced life support (ALS) units staffed by career paramedics. BLS and ALS units are dispatched simultaneously to high acuity calls, including cardiac arrests. Paramedics operate under statewide protocols for cardiac arrest. After consultation with a physician, paramedics may terminate resuscitation attempts on-scene after 20-30 min of unsuccessful efforts.
Data sources and record linkage
The study data consisted of three statewide administrative databases containing patient-level EMS, hospital utilization, and death information, which we supplemented with hospital-level survey data to identify TH centers. We obtained EMS data from the NJ EMS Data Warehouse (EMSDW), which was created as a result of a statewide EMS review mandated by the NJ state legislature and allows EMS practitioners to record all clinical and demographic data on electronic health records [13]. The EMSDW consolidates these electronic records into a single statewide data set. All ALS agencies are required to submit data to the EMSDW. While BLS agency participation is not required, approximately 50 % provide data to the EMSDW. The data captured by the EMSDW include vital signs, procedures performed, response times, resuscitation attempts, patient demographics, and patient identifiers.
The source of hospital utilization data is the New Jersey Discharge Data Collection System (NJDDCS), which contains the universe of uniform billing (UB) records for all inpatient and emergency department (ED) encounters in the state’s hospitals. Hospitals submit claims on a daily basis to the NJ DOH, which edits and standardized claims and retains them in a centralized database. The data captured by the NJDDCS include diagnoses, procedures, patient demographics, and discharge disposition.
The source of mortality data is the state vital records system maintained by the NJ DOH. Under agreements with other states, the NJ DOH obtains mortality records for NJ residents who died outside of NJ.
We linked the data sets using LinkKing© software, which employs a combination of probabilistic and deterministic linkage methods [14]. Linkage was based primarily on patient name, date of birth (DOB), and Social Security Number (SSN) with patient sex, race, ethnicity, and residential zip code as additional linking variables. Although name, DOB, and SSN are not available on public use files, they are maintained by the NJ DOH. Under a special arrangement for this project, the Department’s Center for Health Statistics implemented the required linkages using patient identifiers under state auspices, and delivered a de-identified, linked database to the study team for further preparation and analysis.
Patients
We included adult (ages 19 and older) patients who were treated by EMS for OHCA. Using the EMSDW, we identified cardiac arrests as individuals coded as “cardiac arrest” for call type, those receiving CPR or defibrillation, and individuals with a first monitored cardiac rhythm of ventricular fibrillation (VF), ventricular tachycardia (VT), pulseless electrical activity (PEA), or asystole. We excluded from our analysis all cases where resuscitation attempts were terminated in the field, those where the EMS record could not be linked to a hospital record, and hospital transfers where patients could not be followed throughout the entire episode of care.
Outcome measures
We examined two outcomes: (1) neurologically intact survival to hospital discharge and (2) neurologically intact 30-day survival (i.e., 30 days after the cardiac arrest). Following the approach used in the clinical trial conducted by Bernard et al. [7], we used hospital discharge codes in the NJDDCS to proxy neurological status (and mortality records to measure survival within 30 days of the arrest). Specifically, we defined neurologically intact survival as discharge/transfer to home/self-care, another hospital or short term acute care facility, rehabilitation facility, court/law enforcement, or left against medical advice. We defined poor neurological outcomes as all other discharge destinations (e.g., discharged to hospice or nursing home) and death.
Identification of TH centers
Since the application of TH usually does not affect hospital reimbursement, this procedure is rarely recorded in hospital billing records. Thus, we were unable to utilize hospital billing records to identify individual patients who received TH. Instead, we used a prior survey of acute care hospitals in NJ to define TH centers as hospitals that adopted TH in the treatment of initial OHCA survivors [9]. Since some hospitals initiated their TH programs during the study period (2009-2010), we classified these facilities as TH centers during and after the month of implementation and as non-TH centers in the prior months. Among the state’s 73 full-service hospitals, 18 adopted TH before the study period began, 23 adopted during the study period, and 32 did not use the procedure at all.
Analysis
We estimated the association between treatment at a TH center and the two OHCA survival outcomes using multiple logistic regression. Model covariates included incident, patient, and hospital characteristics. The incident characteristics we considered were year, whether the arrest was witnessed before EMS arrival, defibrillation attempted, shockable rhythm, response time, scene time, and transport time. The patient characteristics we considered were sex, age, race/ethnicity, and insurance coverage (defined as expected primary payer on the hospital record). The hospital characteristics we considered were number of beds, membership in the Council of Teaching Hospitals (COTH), and the poverty rate in the hospital service area (defined as the smallest number of zip codes accounting for at least 90 % of all hospital volume).
As mentioned above, patient transportation to TH centers is potentially endogenous, since prehospital EMS personnel may exercise discretion in their choice of destination hospital. In theory, prehospital EMS personnel may be more likely to take patients to TH centers when they believe these patients are good candidates for TH treatment due to patient risk factors that are unobservable in our data. If so, then ordinary/naive estimates of the relationship between transportation to a TH center and survival outcomes would be biased by unmeasured confounding [15]. To address this issue, we constructed an instrumental variable (IV) to separate patient-level variation in TH versus non-TH hospital into two components: 1) an exogenous component unrelated to likelihood of survival and 2) the remaining component, which contains unmeasured and potentially endogenous patient characteristics that may be related to likelihood of survival.
In this context, a suitable IV must be strongly related to whether a patient is transported to a TH center but have no direct relationship to OHCA survival [16]. In other words, the IV affects survival only through its association with transport to a TH center. For our analysis, we used an IV defined as the differential distance between the closest TH and non-TH hospitals. The IV calculation was based on the distance between the geographic centroid of the EMS incident zip code and the street addresses of the nearest TH and non-TH hospitals.
Since we had exactly one IV to predict exactly one potentially endogenous treatment variable, our data did not satisfy the over-identification condition required to formally test whether our differential distance variable is exogenous (i.e., not directly related to survival outcomes) [15]. Nevertheless, distance-to-hospital variables such as ours are often used as IVs on the grounds that direct association with patient outcomes is generally considered implausible [17–19].
We implemented our IV strategy using the 2-stage residual inclusion method developed by Terza et al. [20], which extends previously developed IV estimation for linear models to non-linear models such as logistic regression. In stage 1, we used ordinary linear regression to predict transport to a TH center based on the IV and all other exogenous independent variables listed above. In stage 2, we estimated logistic regression models to predict each survival outcome (i.e., in-hospital and 30-day) on the basis of transport to a TH facility, the residual from stage 1, and the exogenous independent variables. After controlling for these factors, the coefficient for the TH variable provides a consistent estimate of the relationship between TH transport and survival outcomes by Terza et al. [20].
Additionally, the coefficient for the stage 1 residual in the stage 2 model provides a test of the endogeneity of transportation to a TH center. If this coefficient is significantly different from 0, endogeneity is confirmed. Otherwise, more efficient and consistent estimates are generated from logistic regression without the stage 1 residual [15, 20].
We used the Stock-Yogo F-test to confirm that our IV is sufficiently strong to reliably account for unmeasured confounding [16]. We also used likelihood ratio tests to choose between ordinary logistic regression and mixed effects logistic regression with hospitals specified as random clustering units. All analyses were performed using STATA 13.0.