The Georgia Department of Public Health Office of Vital Records is responsible for collecting information about deaths among Georgians by using the death certificates. The death certificate contains information on individuals’ demographic characteristics, residence, underlying possible causes of death, location of death, and death date. Each year, more than 67,000 Georgians die, and 98% of the deaths occur within the state of Georgia.

The GDDS is housed at the Georgia Hospital Association and has information on all inpatients discharged from nonfederal short-stay hospitals in Georgia. GDDS gathers more than a million records per year. GDDS and mortality data share common variables including age, sex, race, residence information, and a quasi-unique subject identifier (LONGID) that facilitates the data linkage.

The feasibility of data linkage is based on the assumption that the variable LONGID was sufficiently specific to distinguish each subject in the data sources. The LONGID is a 15-digit alphanumeric unique code created from letters of patients’ first and last names, birth date, and sex. We tested accuracy of data linkage by using data from GDDS for 1,494 Georgia patients who were admitted to a hospital for acute stroke and who died as a result (International Classification of Diseases, Ninth Revision [ICD-9]) codes 430–436) in 2006 and for 3,598 patients with similar age characteristics (patients with malignant neoplasm of respiratory and intrathoracic organs: ICD-9 codes 160–165) but who were alive in 2006. Patients with similar age characteristics were chosen because personal name patterns in a given community may change through time.

The test data set was then linked with the Georgia mortality data for 2006 by using a multistage deterministic and probabilistic matching algorithm and various combinations of key variables (Table 1 and Table 2). We used fine-grained record integration and linkage software for matching, and we excluded duplicate entries using the LONGID, admission and discharge dates, and facility codes (6). Degrees of linkage between hospital discharge data and mortality data were determined (Table 3).

We used the 2005 through 2009 GDDS and the 2006 through 2010 Georgia Office of Vital Records mortality data to examine the survival rates of acute ischemic stroke patients. Patients admitted to nonfederal acute care and critical access facilities with the principal diagnosis ICD-9 codes 433 and 434 were identified and linked to the death data. Death and survival time from the index admission date, regardless of the underlying cause of death, were the outcome variables. We believe that care in the first few hours after stroke symptom onset determines the stroke patient’s subsequent health condition, so we attributed the outcome to the facility of first stroke admission. Patients were labeled to have had a first stroke admission in 2006 if they were not admitted for any type of stroke, ICD-9 codes 430–438, in 2005.

We defined enrollment in GCASR if the hospital actively participated in data entry and quality improvement activities. We considered patients to have had stroke care by a GCASR facility if the hospital in which patients were admitted was enrolled in the registry. Patients who were treated at any time before a facility was enrolled or after it withdrew its participation were counted as patients treated by a non-GCASR hospital. We included patient’s sociodemographic characteristics such as age, sex, race, insurance status, and length of hospital stay, and hospital features including number of beds and location as covariates in the analysis. On the basis of the number of beds, we classified hospitals as small (<100 beds), medium-small (100–249 beds), medium-large (250–399 beds) and large hospitals (≥400 beds). We used the Rural-Urban Commuting Area classification of location to classify hospitals geographically into metropolitan (codes 1–3) and nonmetropolitan (codes >3) (7).

Comorbidities were included in the analyses to adjust for disease severity. We used the Healthcare Cost and Utilization Project software from the Agency for Healthcare Research and Quality (US Department of Health and Human Services) to define comorbidities for each patient based on the ICD-9 codes in the hospital discharge data (8,9). Patients’ readmission status before either the end of the follow-up period or the patient’s death was captured from the hospital discharge data, and we classified patients as either not having been readmitted, readmitted to the same hospital, or readmitted to a different hospital. If patients were admitted to a different hospital within a day after their index or first admission date we considered their status as a transfer rather than a readmission, and they were excluded from the analysis. All the variables used in the analyses refer to what was documented at the first stroke admission except for the date of death. To have stable estimates, we excluded stroke patients from hospitals with fewer than 15 patients over the study period.

We analyzed the data by using SAS for Windows (version 9.3, SAS Institute, Inc). We assessed the sensitivity of the linkage procedure based on the proportion of stroke-related in-hospital deaths that were captured by the 2006 Georgia vital records mortality data. We determined specificity by the proportion of subjects who were admitted in 2007 having a malignant neoplasm of the respiratory organs that were linked to any of the records in the 2006 death file. We assessed patient and hospital characteristics descriptively and tested differences between patients treated at GCASR participating and nonparticipating hospitals using χ² tests for nominal variables and Wilcoxon tests for quantitative variables.

We assessed the proportional hazard assumption graphically and through the goodness-of-fit test for correlation between the Schoenfield residuals and failure time (10). We repeated the graphic assessment using the log–negative log of survival curves after adjusting for covariates. The GCASR participation variable did not satisfy the proportional hazard assumption. Thus, we analyzed survival time in correlated data using the extended Cox proportional hazard model with the robust sandwich estimate option to estimate the marginal covariate effects. We performed the analysis with and without censoring at 1 year. Results are presented indicating the hazard ratio for death in the first year after the seventh day of the first stroke admission date by different patient and hospital characteristics, including participation in GCASR.

Of the 1,494 acute stroke patients with an in-hospital death recorded in the 2006 hospital discharge data, 1,381 (92.4%) were identified in the 2006 death data, whereas none of the 3,598 patients with malignant neoplasm of respiratory and intra-thoracic organs diagnosed in 2007 were linked to the 2006 death data. Agreements between hospital discharge records and death data were high (>91%) for demographic variables, facility (93.6%), and discharge or death dates (92.6%) (Table 3).

From the initial 50,937 patients listed, 358 were excluded because 269 were considered transfers and 89 were from hospitals with fewer than 15 cases. Analysis was performed for 50,579 acute ischemic stroke patients (Table 4) admitted to 131 acute care and critical access hospitals in Georgia to assess the impact of participation in GCASR during 2006 to 2009. Most (52%) were women, and whites accounted for two-thirds (66%) of the patients. The mean age for first stroke admission was 69 years. Most (64%) had Medicare as their principal health insurance coverage. The median hospital length of stay was 3 days.

GCASR-participating hospitals treated 56% of the ischemic stroke patients (n = 28,077), and there were no statistical differences in age, hospital length of stay, proportion of various racial groups, or proportion of subjects with insurance coverage between patients treated at GCASR and non-GCASR hospitals (Table 4). However, non-GCASR hospitals were more likely to see female stroke patients, have less than 100 beds, to be in nonmetropolitan areas, and record more stroke-related deaths at 30 and 365 days following stroke admissions. The overall mortality at 30 days and 365 days after the first admission were 8.1% and 18.5%, respectively.

The extended Cox model indicated that patients treated at non-GCASR hospitals had a hazard ratio of 1.14 (95% confidence interval [CI], 1.03–1.26) from the eighth day after admission to 1 year after admission (Table 5). A similar hazard ratio (1.13; 95% CI, 1.04–1.22) was observed when no cutoff date was applied. Similarly, older patients and those treated in nonmetropolitan hospitals had a higher hazard ratio than their counterparts. Patients with a private insurance or self-pay had a lower hazard ratio than did Medicare patients. In addition, hospitals with fewer than 100 beds and longer hospital stays for patients were independently associated with subsequent death (Table 5).