All fee-for-service Medicare beneficiaries aged ≥65 years discharged with an incident ischemic stroke (primary diagnosis: ICD-9-CM 433, 434 and 436) from January 1, 1994, to December 31, 1996, January 1, 1997, to December 31, 1999, and January 1, 2000, to December 31, 2002, were identified from Medicare Provider Analysis and Review files of the Centers for Medicare and Medicaid Services. Incident stroke patients were defined as those who had not had an ischemic stroke hospitalization within the year prior to their index admission. The records were cross-linked with the Medicare Enrollment Database (denominator file) and the patients grouped by their county of residence.
Beneficiaries were excluded if they transferred between acute care facilities, were discharged from a nonacute care hospital, were discharged within 1 day (as it is unlikely that these patients suffered a stroke), were discharged against medical advice, died during the index stroke hospitalization, had crossovers in coverage between fee-for-service and health maintenance organization status or had <12 months of continuous Medicare fee-for-service enrollment. Patients were excluded if their county of residence was unknown, if they lived outside the USA or if had missing follow-up information.
Age, sex, race/ethnicity and ZIP code of residence were obtained from the Medicare Enrollment Database. Prior diagnoses including acute myocardial infarction (AMI; ICD-9: 410), cancer (ICD-9: 140–172.9, 174–195.8) and congestive heart failure (ICD-9: 428–428.9) as well as comorbidities [diabetes (ICD-9: 250–250.3, 250.7), hypertension (ICD-9: 401), dementia (ICD-9: 290–290.9), chronic obstructive pulmonary disease (ICD-9: 490–496), obesity (ICD-9: 278.00) and current smoker (ICD-9: 305.1)] and the Deyo comorbidity index [7
] were determined using ICD-9-CM diagnostic codes from both the index stroke admission and hospitalizations during the prior 12 months. The number of hospitalizations in the year prior to the index stroke was dichotomized as <2 versus ≥2.
Recurrent ischemic stroke hospitalizations were identified by primary diagnosis ICD-9 codes 433, 434 and 436 for 12 months following the index stroke admission. Although there is no standard definition of a recurrent ischemic stroke [8
], prior cohort studies have excluded events within the first 24 h to 28 days [8
]. For this analysis, ischemic stroke hospitalizations occurring within 7 days after the discharge date were classified as complications of the incident stroke and were not considered a recurrent event. The Death Master File was used to determine patients’ survival or date of death as it contains 95% of all deaths among individuals aged 65 years and older [11
The patient's county of residence was identified using the Federal Information Processing Standards (FIPS) code. The ZIP codes were mapped to the county by the Federal Information Processing Standard code that had the greatest proportion of the ZIP code's population, using data from the census. We estimated that 96.2% of all patients would be assigned the correct county by summing the proportion of each ZIP code's population in the primary county, similar to the methods used in other studies [12
Recurrent ischemic stroke rates were calculated as the number of patients with a recurrent ischemic stroke per county divided by the number of patient days at risk for each time period. The percent change in risk-standardized 1-year recurrent stroke rates for each county was calculated as the difference between the rate for 1994–1996 and 2000–2002. Patient characteristics were compared across cohorts using χ2
and t tests as appropriate. County-specific recurrence rates and changes in recurrence rates were mapped to provide a detailed picture of recurrent stroke trends (fig. , ). In order to quantify regional differences, counties were grouped into 9 regions based on US Census Divisions (online suppl. fig. 1
]. The mean risk-standardized recurrent stroke rates and the proportion of counties above the national average were compared by region over time.
Counties with 1-year recurrent stroke rates above the national mean from 1994 to 2002 (shown in black); 23 counties were excluded due to percentage errors of >20% for risk-standardized rates at any of the time periods.
Percentage change in risk-standardized 1-year recurrent stroke rates by US county from 1994–1996 to 2000–2002; 23 counties were excluded due to percentage errors of >20% for risk-standardized rates at any of the time periods.
A spatiotemporal bayesian Poisson conditionally autoregressive model [14
] was used to determine the risk-standardized rates of stroke recurrence for each county. The conditionally autoregressive model smoothed county risk estimates toward the mean risk of the adjacent counties to provide more precise and reliable estimates, especially for counties with small sample sizes [15
]. In addition, this model smoothed rates for each county across adjacent time periods, assuming that changes in rates across counties and over time occur without sudden great variations (i.e. great variations were considered sampling errors and were not included in the final figures). Patient characteristics were included as fixed effects and standardized as the difference from the national mean, providing risk-standardized 1-year recurrent stroke rates for each county. These rates were adjusted for county differences in patient demographics including age, medical history and the prevalence of comorbid conditions.
Model parameters were estimated using Markov chain Monte Carlo simulation, a sampling algorithm used to estimate model parameters in bayesian analyses commonly used for small-area analyses in which rates are determined for relatively small, less populated geographic areas [16
]. An uncorrelated (heterogeneity) random effect was incorporated into the model to control for any extra-Poisson variation due to omitted variables, and the deviance information criterion [17
] was used to determine model goodness of fit.
The models were run on 3 parallel Gibbs sampler chains and monitored using sample autocorrelations within the chains, plots of sample traces and plots of the Gelman-Rubin reduction factor [18
]. Each chain was run for 50,000 iterations, with the first 20,000 iterations being discarded as preconvergence burn-in, and every 10th iteration was retained to reduce autocorrelation. Relative risk estimates and 95% credibility intervals were determined for all covariates. All bayesian modeling was performed with the 64-bit version of WinBUGS 1.4.3 [19
]. The risk-standardized 1-year recurrent ischemic stroke rates were mapped by county and classified by quintile. All maps were created using ArcGIS 9.2 by Esri.