Data source and sample
We analyzed the responses of 4404 adults (≥18 years) with kidney disease from the United States using the Medical Expenditure Panel Survey (MEPS) for the years 2002–2011. MEPS is a nationally representative survey of the U.S. civilian non-institutionalized population and is administered by the Agency for Healthcare Research and Quality [19, 20]. MEPS obtains comprehensive information on participants’ use of medical care, prescription medication and their medical spending, as well as information on demographics, socioeconomics and satisfaction with health care. The complex survey design includes multistage sampling, clustering and stratification with oversampling of minorities [21]. Using weights provided by the Agency for Healthcare Research and Quality to account for sampling, the weighted sample for this analysis represented 4,251,129 adults with kidney disease living in the United States.
Self-reported information is collected from respondents, in addition to collection of data on medical and financial variables from all types of health care providers for validation and supplementation [19]. Diagnosis coded according to International Classification of Disease, Ninth Revision, Clinical Modification (ICD-9-CM) are also collected. Kidney disease related medical conditions and procedures reported by respondents were recorded by an interviewer as verbatim and then converted by professional coders to ICD-9-CM codes. Fully specified ICD-9-CM codes were collapsed into three digits in order to protect confidentiality of respondents [20]. The error rate for any coder did not exceed 2.5 % on verification [22]. For each year, data were merged from the medical condition files and the full-year consolidated files using the unique person identifier (DUPERSID) on a one-to-one match [20]. To ensure sufficient sample size and robust estimation for our analysis [23–25], we pooled the 10-year MEPS data.
Measures
All measures are based on previously validated questionnaires that are publicly available on the MEPS website [19, 20]. Individuals with kidney disease were identified from the MEPS household medical condition files using clinical classification categories (CCCs) codes of 156 (nephritis, nephrosis, renal sclerosis), 157 (acute and unspecified renal failure), 158 (chronic renal failure),160 (calculus of urinary tract) and 161 (other diseases of kidney and ureters) [20].
The dependent variables in this study were dichotomized variables created from questions asked in the access section of MEPS. The first access variable was Usual Source of Care, as determined by response to the question: Do you have a usual source of care provider? An overall Medical Access to Care variable was created from responses to three separate questions: 1) Do you have a usual source of care provider?, 2) Were you unable to get necessary medical care?, and 3) Were you delayed in getting necessary medical care?. If respondents answered ‘yes’ to having a usual source of care, ‘no’ to being unable to get necessary care, and ‘no’ to having a delay in necessary medical care they were coded as having Medical Access. Opposite answers to any of the three questions resulted in being coded as not having Medical Access. An overall Prescription Access to Care variable was created from responses to two questions: 1) Were you unable to get necessary prescription medication?, and 2) Were you delayed in getting necessary prescription medication? [20]. Similarly to Medical Access, if respondents answered ‘no’ to being unable to get prescriptions, and ‘no’ to having a delay in necessary prescriptions they were coded as having Prescription Access. Opposite answers to either of the two questions resulted in being coded as not having Prescription Access.
The primary independent variable was census region coded as: Northeast, Midwest, South and West.
Covariates
All covariates used for analysis were based on self-report and were included to take into account sociodemographic differences between the regions. Binary indicators of co-morbidities were based on a positive response to a question “Have you ever been diagnosed with…?”. Cardiovascular disease (CVD), however, was a positive response to diagnosis with coronary heart disease, angina, myocardial infarction, or other heart diseases. Race/ethnic groups are categorized as: Non-Hispanic White (NHW), Non-Hispanic Black (NHB), Hispanic or others. Education was categorized as: less than high school (≤ grade 11), high school (grade 12) and college or more (grade ≥ 13). Marital status was coded as: married, non-married (widowed/divorced/separated) and never married. Gender was dichotomized and age was coded into three age groups: 18–44, 45–64 and ≥ 65 years. Metropolitan Statistical Areas (MSA) was dichotomized based on population as of end of the year. Health insurance was coded as: private, public only and uninsured at all time in the year. The income level was defined as a percentage of the poverty level and grouped in to four categories: poor (<125 % of poverty level), low income (125 % to less than 200 % of poverty level), middle income (200 % to less than 400 % of poverty level) and high income (≥400 % of poverty level). Calendar year was grouped into 2002/03, 2004/05, 2006/07, 2008/09, 2010/11 for the pooled data.
Statistical analysis
We used multiple logistic regressions with binary variables of Usual Source of Care, Medical Access to Care, and Prescription Access to Care as the dependent variables (yes versus no) across geographic region, adjusting for age, sex, race/ethnicity, marital status, education, insurance status, MSA status, household income, comorbidities and calendar year. For interpretation, we use the adjusted odds ratio coefficient of the logistic regression.
F-adjusted mean residual goodness-of-fit was applied to test the adequacy of the models. After fitting the logistic regression models taking the survey design in to account, the F-adjusted mean residual goodness-of-fit suggested no evidence of lack of fit [26]. The link test that account complex survey design, used as a diagnostic test to examine the model specification error, verified no evidence of model specification error in the models [27]. Using the Variance inflation factor (VIF) test, and taking into account the complex survey design, it was determined that no multicollinearity problems existed between predictors of the models. All analyses were performed at the person-level using STATA 14 (StataCorp LP College Station, TX). Only estimates that are statistically significant at the p < 0.05 level are discussed.