This study used a combination of patient- and practice-level data to create a hierarchical dataset of North Carolina State Employees Health Plan (SHP) diabetes patients nested within North Carolina primary care practices over time. This data set linked Community Care of North Carolina (CCNC) practice participation to patient-level health and demographic information for years 2002, 2005-2008.
Data Sources
Patients’ quality of care, health care utilization, expenditures, and demographic measures were derived from SHP claims data for years 2002, 2005-2008. The SHP data included every health service reimbursed by the SHP including inpatient care, outpatient care, physician care, emergency department use, and prescription medications (Carolina Cost and Quality Initiative, n.d.). The data also included diagnoses, expenditures, and patient descriptors. The SHP claims also contained data on servicing provider, payment provider, provider type/specialty, and provider county for each outpatient physician claim and were linked to SHP administrative information on practice name, recorded date of practice enrollment in SHP, recorded date of disenrollment in SHP, and National Provider Identifier (NPI)
number. The SHP data consisted of three separate files: 1) the member file, which included information on enrollment and disenrollment dates, Medicare start date, birth date, death date, etc.; 2) the
combined claims file, which included all health care utilization claims; and 3) the pharmacy file, which included all pharmacy claims.
For practices that ever participated in CCNC (years 1998-2014), CCNC administrative data contained practice name, practice location address, CCNC network, date practice enrolled in CCNC, date practice disenrolled in CCNC, number of unique CCNC patients assigned to a practice in each year, practice type/specialty, and National Provider Identifier (NPI) number.
Sample
Patients
The sample for this study included SHP beneficiaries who were age 18 and older, diagnosed with diabetes, and had at least one visit to a primary care practice (in order to attribute to a primary care practice) during the study period. Using ICD-9 diagnosis codes (250.XX, 357.2X, 362.0X, 366.41, 648.0X), individuals were identified as having a diagnosis of diabetes if they met any of the following criteria: 1) at least one face-to-face acute inpatient or emergency department (ED) visit with a diagnosis of diabetes (acute inpatient visits were identified as those with an inpatient claim type or a claim with an inpatient hospital place of service code during the days of a hospital admission; ED visits were identified as visits with an ED indicator), 2) at least two face-to-face outpatient or nonacute inpatient visits on different dates of service with a diabetes diagnosis (outpatient CPT codes: 92002, 92004, 92012, 92014, 99201- 99205, 99211-99215, 99217-99220, 99241-99245, 99341-99345, 99347-99350, 99384-99387, 99394- 99397, 99401-99404, 99411, 99412, 99420, 99429, 99455, 99456; nonacute inpatient CPT codes: 99301- 99313, 99315, 99316, 99318, 99321-99328, 99331-99337), or 3) at least one insulin or oral
hypoglycemic/antihyperglycemic prescription event (HEDIS, 2008). This was based on the HEDIS criteria for identifying patients with diabetes (HEDIS, 2008); however, I expanded the IP and ED criteria in order to identify patients with diagnoses on either physician or facility claims. The HEDIS-recommended OP and non-acute IP CPT codes used were also somewhat restrictive, though I only excluded approximately
period were excluded to eliminate the possibility of misdiagnosis of gestational diabetes; these women were identified using a combination of ICD-9 and CPT codes for births, either live or otherwise.
Patient-months were excluded if a patient was enrolled in Medicare (due to incomplete health services, pharmacy utilization, and cost data in SHP claims) or was retired (due to potential differences compared to the working age population, e.g., retired early due to illness). Observations before the first diagnosis of diabetes were excluded. Patient-years were excluded if a patient had less than 10 months in the SHP in that year (only months with full enrollment were included). Ten months was considered a reasonable length of time in which to accrue these outcomes at the person-year level (mean # months in SHP per year=9.5). Enrollment dates for SHP patients occurred at all times during the year, so the 10 month requirement did not inadvertently require two years of enrollment, which could have occurred if annual enrollment was at the same time every year (e.g., July 1). In fact, of the 18,461 individuals that never had a year with >=10 months in the SHP, 80% of these also never had >=10 months in the SHP overall during my study period. Individuals without a year with >=10 months in the SHP were generally older than the study sample and more were male, they had fewer outpatient, inpatient, and emergency department visits. Patient-years before the first visit to a primary care practice were also excluded as patients could not be attributed to a practice in years before the first visit. The initial patient sample for this study was 29,694 patients (Figure 3.1).
Technical Note on Recovering Missing Data
During the course of data cleaning I discovered that the enrollment dates in the member file were not always correct (i.e., some patients had claims dated earlier than the enrollment date) (Figure 3.2). Complete enrollment data was important because the enrollment file contained the Medicare start dates and SHP enrollment dates that I used to identify exclusion criteria (e.g., I required 10 months in
the SHP in each year for my sample and I could inadvertently exclude some person-years due to incorrect enrollment dates).
Following discussions with the Carolina Cost and Quality Initiative (CCQI), the holders of the SHP data, I was granted access to the raw member data. This raw data, as well as the combined claims and pharmacy data, contained two different individual identifiers: 1) individualid is unique to a patient, and 2) newid is not unique to a patient and typically changes after any breaks in SHP coverage. For example, if a patient was enrolled in the SHP from 2002-2004 and then from 2006-2008 he/she should have one individualid and two newid numbers. The raw data set was structured so that patients could have multiple rows of member information; the patient had a new row, and thus new
enrollment/disenrollment date, each time they changed insurance product and/or enrolled in the SHP. Therefore, in order to get a complete view of a patient’s time in the SHP, all rows for an individualid needed to be combined. However, a large number of rows in the raw file had missing individualids with only newids to identify them. Because the individualid is the only person identifier unique to a patient, it was necessary to recover the individualid in order to combine rows and get a complete view of a
patients’ time the SHP. To recover individualids I first merged the rows that had missing individualids with the combined claims using the newid in order to recover the individualid for the enrollment file (Figure 3.3). Through this process I was able to update 165,237 enrollment file rows with recovered individualids. All the recovered individualids were duplicates of those already in the member file. In other words, by recovering the individualids from the claims I was able to recover 165,237 previous SHP spans for the 992,782 patients in the original member file with individualids.
Practices
identifying practices in the SHP claims because multiple servicing providers bill under one payment provider number. For each outpatient physician claim the servicing provider was the individual provider seen and the payment provider was the practice in which the patient was seen.
Identifying Primary Care Practices
To create an initial set of all primary care practices in which any SHP patient (either with or without diabetes) was seen during the study period, I first utilized SHP claims data to create a list of all payment provider identification numbers recorded in the claims during that time (N=52,510); the top specialties: 33% out of state, 19% multispecialty, 4.7% chiropractic, 4.5% family medicine, and 3% internal medicine). I then restricted the sample to practices (i.e., payment provider id numbers) with a primary care specialty. I defined primary care specialty as having a specialty code for family medicine, general medicine, internal medicine, pediatrics, public health, rural health clinic, or multispecialty with at least one primary care provider (i.e., a servicing provider with a specialty of family medicine, general medicine, internal medicine, pediatrics, physician assistant, or family nurse practitioner). I required at least one primary care provider for multispecialty practices in order to ensure at least some primary care focus. I linked this sample to the SHP administrative information on practice name, date of practice enrollment in SHP, date of disenrollment in SHP, and National Provider Identifier (NPI) number. Using this method I identified an initial sample of 5,851 primary care practices. Of these practices, 51% were family practice, 30% were internal medicine, 9% were multispecialty, 6% were pediatrics, 4% were general practice, and 2% were public health. Distributed among these practices (i.e., payment provider ids) were 18,872 primary care providers (i.e., servicing provider ids) (Figure 3.4).
Attribution of Patients to a Primary Care Practice
SHP patients are not administratively linked to a primary care practice as their appointed usual source of care. Therefore, in order to determine if a patient was in a CCNC or non-CCNC practice, I attributed a patient to a SHP primary care practice (i.e., payment provider number) in each year. Patients were attributed to a primary care practice based on a plurality of office or other outpatient Evaluation and Management codes (CPT codes: 99201-99205, 99211-99215, 99241-99245) to that practice during the measurement year. Visits were defined as a unique combination of patient, provider, and date of service. If a patient had the same number of visits to multiple practices, the patient was attributed to the practice that had the last visit during the year (Center for Medicare and Medicaid Services, n.d.; The Brookings Institution and Dartmouth Institute, 2011). Using this method, SHP diabetes patients were attributed to a total of 2,294 primary care practices out of the initial 5,851 primary care practices. While I attributed adult diabetes patients to under half of the total 5,851 primary care practices, further examination of the data indicated that many of the 5,851 practices were either out-of-state, pediatric practices, or in my study period for only one or two years.
Identifying CCNC Practices
I identified 2,294 primary care practices in which SHP diabetes patients were attributed. To determine which practices participated in CCNC, I first linked the SHP primary care practice sample with the CCNC administrative data using NPI. Of the 2,294 SHP primary care practices 948 practices matched to a CCNC record via NPI (Figure 3.5). Because the mandated use of NPI was only implemented in 2007 (CMS, 2007), practices that closed before 2008 did not have an NPI number (N=296). Additionally some practices changed NPI number over time, thereby having multiple NPIs, which resulted in the possibility of a non-match in NPI due to different NPIs listed in the SHP or CCNC data. Therefore, for SHP practices
matched CCNC practices using name, county, enrollment date, and disenrollment date. For example, I matched the SHP practice “Ahoskie Medical Associates” in Hertford County, which started in SHP in 1974 and ended in September 2010 to the CCNC practice “Ahoskie Medical Associates” in Hertford County, which disenrolled from CCNC in August 2010. Using this method I was able to match an
additional 107 practices. To verify the hand matches I linked the SHP practices to the NPI database using the NPI in order to recover the business location address of practices in the NPI database. I then
compared the recovered business location address with the location address listed in the CCNC administrative data. I verified a total of 44 practices. Of the remaining 63 practices, 46 closed prior to 2008 and did not have an NPI and so did not have an NPI business address to compare. Through the verification process also I matched an additional 14 practices that I had not previously hand matched, resulting in a total of 121 practices hand-matched.
There were some practices that had multiple identification numbers, either NPI or provider number, in the SHP or CCNC data; therefore the count of 2,294 practices was too large as it counted some practices more than once. To identify practices with multiple entries I used a combination of SHP practice name, enrollment and disenrollment dates, and identification number; multiple entries were considered one practice if they had the same name as well as consecutive disenrollment and enrollment dates (e.g., Carolina Family Prac & Sports Med had a disenrollment date of 05/31/2008 and Carolina Family Practice & Sports M had an enrollment date of 06/01/2008), or same name and it was that of a physician (Figure 3.6). Combining entries with consecutive disenrollment and enrollment dates and same physician name resulted in a practice sample of 2,140 practices. I updated patient attribution so that patients were attributed to one of these final 2,140 practices each year (Figure 3.7), of which 878 (42%) were ever in CCNC during my study period. In 2007 it was estimated that just over 50% of primary care practices in North Carolina were participating in CCNC (Steiner, 2009); because I am excluding many
pediatric practices, due to my study sample only including adults, 42% seems a plausible participation rate for my study sample.
Additionally, in some instances multiple CCNC practices shared the same NPI, which linked to the same SHP practice. For example, the CCNC practices “Park Ridge Medical Associates” and “Laurel Park Medical Center” were both linked to one SHP practice, “Park Ridge Medical Assoc” via NPI. This occurred for a total of 110 CCNC practices, grouped into 26 SHP practices, with the number of CCNC practices per SHP practice ranging from 2-13. I identified the practices at the SHP practice level (i.e., I identified 26 practices) but created indicators to identify that they were a part of a multi-site practice. Twenty-two of the 26 practices had sites with varying CCNC enrollment dates. I created indicators to identify these as mixed-CCNC practices.
Sample Exclusions
The final SHP sample excluded patient-years (~1%) in which a patient was attributed to one of the 54 practices located outside of North Carolina. I allowed individuals to be attributed to these practices, as individuals living near the NC boarder could reasonably seek the majority of their care in another state. However, I excluded these practices from analysis because health care laws and regulations, as well as provider practices can vary by state leading to potential systematic differences between North Carolina practices and those in other states. Additionally I excluded patient-years (~5%) in which a patient was attributed to one of the 22 practices that had multiple practice sites with varying CCNC enrollment dates as the effect of CCNC would be very imprecise and noisy for these practices.
I also excluded patient-years in which a patient was attributed to a practice that disenrolled from SHP before July 1 or joined SHP after July 1 of a year, as these practices were not able to provide a full year of services during these years. Lastly, I excluded patient-years (~1%) in which a patient was
qualified for inclusion in the original practice sample because they had at least one primary care physician during the study period; however, practices did not necessarily have a primary care physician in each study year. As this study focuses on primary care practices, I did not include practice-years in which a practice had no primary care physicians because they would not be considered primary care practices in those years.
With these exclusions, the final patient sample was 28,293 nested within 2,038 practices, of which 858 were ever CCNC. The average number of patients per practice over the study period was 9.76, ranging from 1 to 345 patients.
County Data
Control data at the county level were obtained from several sources including the U.S. Census Bureau, the North Carolina Division of Medical Assistance (DMA), North Carolina Health Professions Data System (HPDS), and the USDA Economic Research Service. The U.S. Census Bureau holds the Small Area Income and Poverty Estimates, which includes the county population below the poverty line, and Intercensal Estimates of county racial/ethnic composition. The DMA holds data on the number of Medicaid patients in each county in North Carolina over the study period. The HPDS contains data on North Carolina county provider composition over the study period, including the number of total physicians and primary care physicians per county. The USDA Economic Research Service contains the 2003 Rural-Urban Continuum Codes used to identify rural/urban status