Data were obtained from the Nationwide Inpatient Sample (NIS), which approximates a stratified 20% sample of all non-Federal, short-term, general, and specialty hospitals serving adults in the United States. Detailed information on the design of the NIS is available at http://www.hcup-us.ahrq.gov
. From 1997 to 2006, NIS captured discharge-level information on several million discharges each year. A unique hospital identifier allows for linkage of discharge data to an NIS dataset with hospital characteristics. The Nebraska and Iowa Hospital Associations (Iowa beginning in data year 2001) prohibit release of discharge data on HIV, and so information on these discharges was therefore not included in the NIS.
All patients with a primary diagnosis of stroke were included whether they were alive or dead at the time of discharge. To analyze percentages of hospital-based stroke diagnoses, all discharges with International Classification of Diseases, Ninth Revision (ICD-9)-CM codes 430.xx–438.xx included in any of the diagnoses were identified. For patients with >1 reported code, only the first reported code was included to avoid double counting patients with multiple reported codes. Total number of stroke hospitalizations was obtained by summing across codes. In a secondary analysis to ensure that all strokes were captured, we included stroke listed as a primary or secondary diagnosis. However, since such an approach could also lead to overascertainment, we utilized an appropriate correction for each code, to account for any potential overascertainment.3
By this procedure, the estimated number of strokes by ICD-9-CM code 430 to 438 was obtained by multiplying the weighted total number of discharges with each ICD-9-CM code by its estimated positive predictive value (PPV) for stroke.3
The PPVs were derived by pooling data from previously published studies.3
We computed the weighted proportion of stroke hospitalizations that occurred in persons who had a comorbid HIV diagnosis across the 10-year study period, by stroke type and overall. Trend p values were computed by including year as a continuous variable in the logistic regression models while adjusting for the NIS survey design. We also plotted the primary stroke hospitalization rates in the general HIV-negative vs HIV-positive population in the United States, and computed the weighted frequency of primary stroke hospitalizations in HIV-positive vs HIV-negative patients in order to obtain the rate numerators. Rate denominators were derived using US census data for the total number of persons with and without HIV in the United States. HIV prevalence data were not available for year 1997, which was therefore excluded from the above plots. Rate comparisons were performed using the Z test.
Univariate logistic regression adjusted for the survey design variables was used to evaluate sociodemographic, hospital-level, and clinical predictors one at a time (unadjusted analysis). To evaluate the independent association of these factors with presence of comorbid HIV diagnosis, we used multivariable logistic regression modeling after adjustment for the survey design variables. In the first multivariable model, sociodemographic (age, race [white, black, other, unknown], primary payer [Medicare, Medicaid, private, other]) and hospital factors (Northeast, Midwest, South, West), bed size (small, medium, large), stroke volume by quartile, and location/teaching status (rural, urban nonteaching, urban teaching) were adjusted for. The second multivariable model adjusted for the variables in model 1 in addition to clinical factors (vascular disease entities including hypertension, myocardial infarction, diabetes, and atrial fibrillation, as well as other chronic general medical conditions including liver disease, renal disease, and chronic pulmonary disease). All data analyses were conducted using SAS (version 9.1; SAS Institute Inc., Cary, NC). To summarize the relation between each of the above factors and the odds of having a comorbid HIV diagnosis, we report the unadjusted and adjusted odds ratios and the corresponding 95% confidence intervals as derived from the above logistic regression models. Statistical hypotheses were tested using p < 0.05 as the level of significance.
Standard protocol approvals, registrations, and patient consents.
Since this was an analysis of a publicly available deidentified administrative online database, formal review by the Institutional Review Board at our institution was not required.