ARIC cohort participants (N=15,792) were enrolled from 1987–1989 from the following four US communities: Forsyth County, North Carolina; Washington County, Maryland; suburbs of Minneapolis, Minnesota and Jackson, Mississippi35
. As part of annual follow-up, information regarding inpatient hospital stays is collected from cohort members, and hospitalization data are abstracted from the medical record.
All-cause hospitalizations are identified during annual follow-up or during routine ARIC community surveillance36
. For the current study, cardiovascular disease (CVD)-related hospitalizations were further identified from all-cause hospitalizations using International Classification of Diseases, Version 9 (ICD-9) discharge codes 402
; while a HF-related hospitalization was defined as that with an ICD-9 discharge code 42837
Participants’ addresses obtained at baseline were assigned to the level of the census tract by a vendor with high geocoding accuracy (Mapping Analytics)38
. The 1990 US census tract-level neighborhood-level socioeconomic measure selected for study was median household income (nINC). In previous work, the use of the single-variable nINC measure produced results of similar magnitude and precision when compared to a more complex composite index measure of neighborhood SES39
. We categorized nINC into community-wide tertiles based upon participants’ place of residence at baseline, during the period 1987–1989: low (<$24,777), medium ($24,777≤–<36,071) and high (≥$36,071).
After excluding 245 participants with prevalent HF at baseline, 1,415 participants had an incident hospitalized HF event through 2004. An additional 70 participants were excluded due to missing data on neighborhood socioeconomic status, and 3 were excluded due to insufficient numbers for analysis because they were not white or black, or were blacks living in Minnesota or Maryland, resulting in a final sample size of 1,342 participants.
Covariates included race/study community, gender, age at incident HF hospitalization and selected socioeconomic, clinical and behavioral characteristics. Educational attainment was assessed at baseline (less than 11 years, high school graduate, and greater than high school), as was health insurance status at the time of the index HF hospitalization (receipt of Medicaid, yes/no). Participants’ body mass index (BMI) was assessed at baseline and classified as normal (<25 kg/m2), overweight (25–<30 kg/m2) or obese (≥30 kg/m2). Hypertensive status at baseline was identified as systolic blood pressure ≥140 mmHg, diastolic blood pressure ≥90 mmHg, or taking hypertensive medication within the previous two weeks. Teaching status of the hospital during the index admission (teaching vs. non-teaching), was based upon whether or not the hospital had an internal medicine residency training program.
We ascertained the prevalence of common underlying conditions at the time of the index HF hospitalization using ICD-9 discharge codes. The Charlson Index, a clinical comorbidity algorithm19
, was derived from these data. The Charlson Index is a validated measure used to quantify the burden of comorbidity in several studies of mortality and adverse health outcomes18,19
. In its use with HF outcomes, a “modified” Charlson Index excludes chronic HF from the conditions included in the computation of the comorbidity score40
. Consistent with previous studies, we defined a high burden of comorbidity as a sum of two or more points on the Charlson Index scale, whereas a low burden of comorbidity was defined with a total of zero to one points.
We used generalized linear Poisson mixed models to estimate all-cause, CVD-related and HF-related rehospitalization rate ratios, comparing the rates of participants from low nINC to high nINC, medium nINC to high nINC and Medicaid recipients to non-Medicaid recipients, along with 95% confidence intervals (RR, 95% CI). This modelling strategy accounted for repeat hospitalizations among patients as well as the clustering of patients within census tracts. Time at risk for rehospitalization was the time elapsed between the incident HF hospitalization admission date and death, loss to follow-up or the end of 2004, whichever came first. We assessed for over-dispersion by consulting the deviance statistic of the Poisson model, and conducted supplementary analyses using negative binomial regression when the deviance statistic exceeded one41
The product-limit (Kaplan-Meier) method was used to measure time to readmission, death, or readmission or death over the course of follow-up. Multivariate Cox proportional hazard models estimated the risk of death or rehospitalization or death, and rehospitalization alone using death during follow-up as the censoring variable. The model produced survival curves depicting survival free of readmission or death, and the proportional hazards assumption was assessed. All participants were censored at the end of 2004.
Crude nINC-rehospitalization/mortality analyses were conducted, the influence of covariates in a full model were tested, and effect modification (pinteraction<0.05) of the nINC-rehospitalization/mortality relationship was assessed by age, race/study community, gender, hypertension, BMI and comorbidity index score. Analyses were performed by using SAS Version 9.1 (SAS Institute, Inc., Cary, NC).