We analyzed data from the Cardiac Arrest Registry to Enhance Survival (CARES), which is a multi-center registry coordinated by the Centers for Disease Control and Prevention and Emory University. Detailed information about this registry, including catchment area, emergency-medical-service (EMS) characteristics, and cardiac-arrest protocols, has been reported previously.^18–21 During the study period, from October 1, 2005, through December 31, 2009, CARES collected data on all 911-activated events involving cardiac arrest that occurred in 29 U.S. sites (Fig. 1S in the Supplementary Appendix, available with the full text of this article at NEJM.org). Within a catchment area of approximately 22 million people, 54 EMS agencies submitted data for all out-of-hospital cardiac arrests. The collection of data on all cardiac arrests by the 911 call center in each city was confirmed during the data-review process. A data analyst employed by CARES validated the data and reviewed every record for completeness and accuracy.²¹ The study was approved by the Emory University institutional review board, which waived the requirement for informed consent because the analysis included only deidentified data.

A total of 20,020 events met the criteria for an out-of-hospital cardiac arrest (see the definition and Fig. 2S in the Supplementary Appendix). We excluded 3682 events (18.4%) that did not meet our eligibility criteria (e.g., the cardiac arrest occurred in a facility with on-site health care professionals, such as a nursing home, hospital, medical clinic, or jail, or occurred in an airport [airports are typically closely monitored and have numerous trained rescuers and publicly accessible defibrillators available]). We further excluded 1883 events (9.4%) that were witnessed by EMS personnel, 82 (0.4%) for which the address at which the cardiac arrest occurred could not be determined, 8 (<0.1%) for which data documenting whether the patient received bystander-initiated CPR were missing, and 140 (0.7%) for which the clinical outcome was missing. Our final cohort comprised 14,225 patients with an out-of-hospital cardiac arrest.

Patient-level characteristics were obtained from the CARES database. Characteristics that were used as predictive variables included age, sex, race or ethnic group (coded by the EMS provider as white, black, Hispanic, other, or unknown), location of cardiac arrest (public vs. private), and whether the arrest was witnessed (by someone other than the first responder or EMS provider).

From CARES we also obtained data for the primary study outcome, which was performance of CPR by a bystander. We defined “bystander” as any person who was not part of the 911 response team. Additional CARES data included whether an automated external defibrillator was used, the cardiac rhythm at presentation, survival to hospital admission, survival to hospital discharge, and neurologic outcome at the time of hospital discharge. Neurologic outcome was coded by the CARES hospital contact with the use of a cerebral-performance category (CPC) scale ranging from 1 to 5, with 1 indicating conscious with normal function or only slight disability, 2 conscious with moderate disability, 3 conscious with severe disability, 4 comatose or in a vegetative state, and 5 brain-dead or dead.^22–24

We geocoded the CARES data set on the basis of the address of the cardiac arrest, using the Centrus Desktop geocoder, version 4.0 (Pitney Bowes). We used census tracts as proxies for neighborhoods, because they represent socially and economically homogeneous groups of approximately 4000 to 7000 people.²⁵ Neighborhood-level variables were linked to each geocoded address with the use of the 2000 U.S. Census Summary Files.²⁶ From this linkage, we identified six neighborhood characteristics on the basis of a priori hypotheses from our previous work¹⁰ as possible predictors of bystander-initiated CPR. These included median age, median household income, percentage of people living below the poverty line, percentage of single-person households, racial or ethnic-group composition, and percentage of people with a high-school diploma or a higher level of education. For both the individual and census-tract characteristics, white race and black race were specified as non-Hispanic white and non-Hispanic black, respectively.

We classified neighborhoods as predominantly white (>80% white) or predominantly black (>80% black). If neither the proportion of black residents nor the proportion of white residents in a neighborhood was more than 80%, we classified the neighborhood as integrated. For our main analyses, we defined low-income and high-income neighborhoods as those census tracts in which the median annual household income was less than $40,000 and $40,000 or more, respectively. We then created a six-category variable to examine the association between the combination of neighborhood racial composition and median income and the provision of bystander-initiated CPR. The categories included low-income black, low-income integrated, low-income white, high-income black, high-income integrated, and high-income white.

The primary outcome for all analyses was performance of bystander-initiated CPR. To determine the associations of individual-level and neighborhood-level characteristics with the performance of bystander-initiated CPR, we used a three-level hierarchical logistic-regression model. This allowed us to account for the nesting of 14,225 patients (level 1) within 2403 neighborhoods defined as census tracts (level 2), and 29 U.S. CARES sites (level 3). Individual-level characteristics (model 1) and neighborhood-level characteristics (model 2) were added to the model as fixed effects and CARES sites were added as random effects in order to examine their independent contributions. The final model was chosen on the basis of the greatest proportion of variance that was explained by individual and neighborhood variables. We then conducted 10-fold cross-validation (i.e., the data were divided into 10 validation subsets) to assess the calibration and discrimination of the model. Finally, we conducted posterior predictions, stratified according to type of cardiac arrest (unwitnessed in a private location [i.e., home], witnessed in a private location, unwitnessed in a public location, or witnessed in a public location), to show the associations between neighborhood and performance of bystander-initiated CPR.

Several sensitivity analyses were conducted to assess the potential effect of different thresholds on the associations in our model (e.g., high vs. low income and black vs. white race). All statistical analyses were conducted with the use of Stata software, version 11.2 (StataCorp). P values are based on a two-sided significance level of 0.05.

Of 14,225 patients with out-of-hospital cardiac arrests that met the criteria for inclusion in the study, 4068 (28.6%) received bystander-initiated CPR. The individual-level and neighborhood-level characteristics of the cardiac arrests that were included in the predictive analyses are shown in Table 1, according to whether bystander-initiated CPR was performed. Patients with cardiac arrest who received bystander-initiated CPR were more likely to be male than female, more likely to be white than another race or ethnic group, more likely to be in a public location than a private location at the time of the arrest, and more likely to have a witnessed arrest than an unwitnessed arrest. Patients were less likely to receive bystander-initiated CPR if they had a cardiac arrest in a neighborhood that was low-income or predominantly black (Table 1 and Fig. 1).

Of the 14,225 patients included in the study, 1144 (8.0%) survived to hospital discharge; 587 patients (4.1%) were discharged home with a CPC score of 1. As compared with patients who did not receive bystander-initiated CPR, those who did were more likely to be found with a shockable rhythm such as ventricular fibrillation or ventricular tachycardia, to receive treatment with an automated external defibrillator, to survive to hospital admission and discharge, and to have a CPC score of 1 or 2 at the time of discharge. Additional characteristics of the study cohort are provided in Table 1S in the Supplementary Appendix.

Table 2 shows the final multivariable model and the adjusted odds ratios and 95% confidence intervals for performance of bystander-initiated CPR. At the individual level, patients who were black or Hispanic were less likely to receive bystander-initiated CPR than those who were white. As compared with persons with cardiac arrest in high-income white neighborhoods, those in low-income black neighborhoods were less likely to receive bystander-initiated CPR. Patients with cardiac arrest in low-income white, low-income integrated, and high-income black neighborhoods were also less likely to receive bystander-initiated CPR. High-income integrated neighborhoods were similar to high-income white neighborhoods with respect to receipt of bystander-initiated CPR. Cross-validation of the final model showed excellent calibration and modest discrimination (Fig. 3AS and 3BS in the Supplementary Appendix). Sensitivity analyses showed that our findings remained unchanged when different thresholds were used to define median income and racial composition of the neighborhoods (Table 2S in the Supplementary Appendix).

The magnitude of census-tract effects on the provision of bystander-initiated CPR stratified according to type of cardiac arrest is shown in Table 3S in the Supplementary Appendix. All other things being equal, patients with a witnessed arrest in a public location in a high-income white neighborhood had a higher probability of receiving bystander-initiated CPR than did patients with a witnessed arrest in a public location in a low-income black neighborhood. There was a similar relationship across all four subgroups of cardiac-arrest types (unwitnessed in a private location, witnessed in a private location, unwitnessed in a public location, and witnessed in a public location); patients with cardiac arrest in a high-income white neighborhood were more likely to receive bystander-initiated CPR than those in a low-income black neighborhood.

Using the estimates from the adjusted model, we determined the probability of receiving bystander-initiated CPR across the six categories of median income and racial composition. Figure 2 shows how the mean probability of receiving bystander-initiated CPR and corresponding 95% confidence intervals differed across these categories.

Because CARES data were submitted by 29 U.S. sites with a combined base population of approximately 22 million, we also wanted to determine how predicted probabilities differed across sites within the same six categories of median income and racial composition (Fig. 4S in the Supplementary Appendix). In 13 of the 14 CARES sites that had low-income black neighborhoods, patients who had cardiac arrest in these types of neighborhoods had the lowest mean predicted probability of receiving bystander-initiated CPR. In 22 of the total 29 CARES sites, patients who had cardiac arrest in high-income white neighborhoods had the highest mean predicted probability of receiving bystander-initiated CPR.