Overview
We conducted a secondary analysis of data from the nationwide REasons for Geographic and Racial Differences in Stroke (REGARDS) study [17]. We used participant characteristics determined during the REGARDS study baseline visit and assessed ambulatory utilization in the 12 months following the baseline visit. Institutional Review Boards (IRBs) of the participating institutions approved the protocol. All participants provided written informed consent.
Setting, population, and data sources
Between 2003 and 2007, 30,239 community-dwelling black and white adults ≥45 years old were enrolled in the REGARDS study [17]. As previously described, the REGARDS sample was selected from commercially available nationwide lists of individuals, divided into strata by geography, race, and sex [17]. The study involved oversampling of participants living in the Southeastern U.S., balanced sampling of black and white individuals, and balanced sampling of men and women [17]. Potential participants were contacted first by mail and then by phone [17]. Exclusion criteria included race other than black or white, Hispanic ethnicity, active treatment for cancer, medical conditions that would preclude long-term participation, cognitive impairment judged by the trained telephone interviewer, residence in or inclusion on a waiting list for a nursing home, and inability to communicate in English [17].
The organizations involved in data collection for the REGARDS study include: an Operations Center and Survey Research Unit (SRU) at the University of Alabama at Birmingham, a central laboratory at the University of Vermont, an electrocardiogram reading center at Wake Forest University, a company that conducted at-home visits (Examination Management Services Inc. [ESMI]), and a medical monitoring and stroke adjudication center at Alabama Neurological Institute, Inc. [17] An Executive Committee consisting of the principal investigator (PI) at each study center, plus a representative from the National Institute of Neurological Disorders and Stroke, assists the REGARDS PI at the University of Alabama at Birmingham with the scientific direction of the REGARDS study [17]. This Executive Committee reviewed all study methods and data collection protocols, which were approved by the participating institutions’ IRBs [17].
Baseline data collection involved computer-assisted telephone interviews and in-home visits with a physical examination, blood test, urine test, electrocardiogram, and medication inventory [17]. This data collection was conducted by approximately 100 trained telephone interviewers and approximately 6500 trained ESMI examiners [17]. Performance by interviewers and examiners was monitored by SRU and ESMI, respectively [17]. Details regarding the data collection instruments, which included many previously validated measures, have been published [17].
For our primary analysis, we used data from the REGARDS baseline in-home study visit. As a sensitivity analysis, we used data from the REGARDS second in-home study visit, which took place between 2013 and 2016, approximately 10 years after the baseline visit, as well as data on adjudicated cardiovascular events that had occurred by the time of the second in-home study visit [18, 19]. For all analyses, we also used participants’ linked Medicare claims for the 12 months after each REGARDS in-home study visit [20].
Variables
Participant characteristics were collected by the REGARDS study and included: demographics (age, sex, race, marital status, annual household income, educational attainment, geographic region, and rural/urban setting), medical conditions (hypertension, dyslipidemia, diabetes, atrial fibrillation, myocardial infarction, and stroke), medications (total number of medications taken in the past 2 weeks; and use of antihypertensive medication, insulin, and/or statin), health behaviors (cigarette smoking status, alcohol use, and exercise frequency), psychosocial variables (being the primary caretaker for another individual, having seen any close friends or relatives in the past month, and depressive symptoms), physiologic variables (body mass index; systolic blood pressure; total, low density lipoprotein and high-density lipoprotein cholesterol; serum glucose; estimated glomerular filtration rate; urinary albumin-to-creatinine ratio; and high sensitivity c-reactive protein), and self-rated health. The definitions of these variables, which incorporate previously validated definitions [21,22,23,24,25,26], can be found in Appendix 1.
We used Medicare claims to identify ambulatory visits for the 12-month period following the REGARDS study visit. Ambulatory visits were defined using a modified National Committee for Quality Assurance (NQCA) definition [27] that was restricted to Clinical Procedure Terminology (CPT) codes for in-person, evaluation-and-management visits for adults in an office setting [7]. The NCQA definition of ambulatory visits does not include emergency department visits. We identified ambulatory providers by considering the Unique Provider Identification Numbers (UPINs) in the claims for the ambulatory visits. For each participant, we determined the percentage of visits with the most frequently seen provider.
For each participant, we calculated a fragmentation score using the Bice-Boxerman Index (BBI) [28], which has been previously validated [9, 12, 29]. This index captures both “dispersion” (the spread of ambulatory visits across providers) and “density” (the relative share of visits by each provider) [30]. Patterns of care characterized by high dispersion (many providers) and low density (a relatively low proportion of visits by each provider) receive worse scores (indicating more fragmentation) than patterns with low dispersion and high density. The original BBI ranges from 0 (each visit with a different provider) to 1 (all visits with same provider). To facilitate interpretation, we reversed the index, calculating 1 minus BBI, so that higher scores reflected more fragmentation [1, 7, 11]. Appendix 2 shows the formula for the BBI. Appendix 3 provides examples of ambulatory care patterns and their corresponding BBI scores.
For additional analyses, we categorized providers as primary care or specialty care, using the National Plan and Provider Enumeration System (NPPES) [31].
Statistical analysis
We included participants ≥65 years old whose REGARDS study data were linked to Medicare claims at baseline. We excluded participants who did not have fee-for-service Medicare, did not have continuous coverage for 12 months following baseline, or died within one year after the baseline visit. We excluded participants who qualified for Medicare on the basis of end-stage renal disease, as utilization patterns for these beneficiaries are substantively different from those of other beneficiaries [32]. We also excluded those who had ≤3 ambulatory visits in the first year of observation, as calculating fragmentation scores with fewer than 4 visits can yield unstable estimates [12]. Finally, we excluded (and then later re-included in a sensitivity analysis) participants who had fragmentation scores that were equal to the ends of the scale (equal to 0.00 or 1.00), as these scores are relatively uncommon and represent ambulatory care patterns that may violate underlying trends [7]. For example, a beneficiary can have a score equal to 1.00 if he or she has 4 visits with 4 different providers, but this is not conceptually “more fragmented” than a beneficiary who has 9 visits with 6 different providers and a fragmentation score of 0.92 [7].
We used descriptive statistics to characterize the final study sample. We compared those included in the study to those who had been excluded on the basis of having ≤3 ambulatory visits, using t-tests, chi-squared tests, and Wilcoxon rank sum tests.
We divided the study sample into quintiles based on their fragmentation scores. We determined the median number of visits, providers, proportions of visits with the most frequently seen provider, and fragmentation scores within each quintile. We calculated p-values for trend across quintiles for each of these measures of ambulatory utilization.
To explore the unadjusted associations of race, income, and education with fragmentation scores, we calculated the percentage of individuals with a given characteristic (black race; annual household income <$35,000; and high school education or less) within each fragmentation quintile. We then calculated p-values for trend across quintiles. To facilitate interpretation, we generated descriptive statistics of ambulatory utilization (visits, providers, percentage of visits with the most frequently seen provider, and fragmentation score) stratified by race, income, and education. We further stratified visits and providers by primary care vs. specialty care. We compared ambulatory utilization patterns across subgroups by race, income, and education, using non-parametric Wilcoxon two-sample tests.
We used Tobit models to determine whether race, income, and education were associated with fragmentation scores. We used Tobit models instead of linear models, because the possible values for fragmentation were bounded. Interpretation of Tobit models is the same as for linear models; coefficients represent the absolute amount of change in the fragmentation score. Because the fragmentation score is on a scale from 0.01 to 0.99, changes in the absolute fragmentation score can be multiplied by 100 to yield an equivalent percent change in the fragmentation score. Unadjusted models considered race, income, and education separately. Model 1 adjusted for race, income and education in the same model. Model 2 added adjustment for age, sex, marital status, geographic region, and rural geography. Model 3 included all variables in Model 2 plus medical conditions and medications. Model 4 included all variables in Model 3 plus health behaviors, psychosocial variables, physiologic variables, and self-rated health.
Analyses were conducted with SAS (version 9.4, Cary, NC). P-values < 0.05 were considered statistically significant.