We employed standard methods of small area analysis to delineate PCSAs (Wennberg and Gittelsohn 1973
). First, a patient origin matrix was computed that tabulated Medicare ambulatory care visits from zip code of beneficiary residence (population zip code) to zip code of generalist provider (provider zip code). Next, to define crude PCSAs, population zip codes were assigned to the provider zip code where the plurality of beneficiaries received care, and these assignments were adjusted to create contiguous zip code groups. The comparison of utilization patterns using pediatric primary claims from Medicaid files in six states and commercial claims in Michigan evaluated the generality of the PCSAs to younger populations. Finally, we characterized the areas in terms of their geography, population, and the degree to which primary care utilization occurred within PCSA boundaries, a phenomenon we call “localization.”
Medicare Claims Data
The principal Medicare data sources for primary care utilization were the 1996 and 1997 Part B 5 percent Physician Provider File and a supplemental file of 3 million additional beneficiaries sampled from low-population zip codes (). This latter file included utilization from all beneficiaries within zip codes having fewer than 140 beneficiaries, the minimum number required for reasonably certain zip code assignment. (Power calculations are available from the authors.) In zip codes with more than 140 beneficiaries, the sampling fraction was adjusted downward to always obtain 140 beneficiaries until the 5 percent file, by itself, sampled 140 or more persons within zip codes. Thus, the 5 percent file was the sole source of Part B data when there were at least 2,800 beneficiaries residing in the zip code. The total number of beneficiaries included in the sample was 2,596,005. Claims were weighted to reflect the aggregate (5 percent plus additional sampled beneficiaries) beneficiary sample of each zip code.
Data Sources for Primary Care Service Area Analyses
Visits at Federally Qualified Health Centers and Rural Health Clinics came from a 1996 20 percent Outpatient File (Welch 1998
) and were weighted accordingly during analyses; 232,400 beneficiaries (unweighted) had claims in the Outpatient File in addition to any Part B utilization.
Zip Code File
We used a 1999 area zip code file (Geographic Data Technologies, Inc., Lebanon, NH) as the geographic building blocks of PCSAs (number of zip codes=30,107). All claims, population, and provider data were mapped to these zip codes for data summaries and analysis.
Patient Origin Study
The following criteria constituted the operational definition of primary care services in the Part B Medicare claims (): ambulatory visits at offices, clinics, and hospital outpatient departments to generalist physicians (family and general practitioners, general internists, and general pediatricians), midlevel providers, and “clinic,” a residual category that constituted primary care as evidenced by diagnostic codes. Emergency room and consultative visits were excluded. Claims within the Outpatient File were restricted to those with provider number and facility type of Rural Health Clinics and Federally Qualified Health Centers that also had a revenue center of “clinic.” These claims constitute primary care patient encounters. The Medicare beneficiary population included those U.S. residents aged 65 and older and who also had “aged” eligibility. The utilization of beneficiaries enrolled in risk-contract HMOs (health maintenance organizations) is not reliably reported, and these patients were excluded (15 percent). Altogether, these criteria resulted in 17,905,758 claims that represented, after sample weighting, 197,797,757 visits.
Criteria for Identifying and Selecting Medicare Primary Care Claims: Part B Files (Selected Claims = 16,852,855)
We used the provider zip code from the claim to ascertain provider location. We identified provider zip codes as those with at least 50 weighted primary care claims. This threshold of claims was intended to reduce the number of zip codes whose apparent primary care activity was from zip code misspecification in the claim form or from extremely part-time provision of care. Population zip codes were those with at least one resident beneficiary.
Each beneficiary's utilization was analyzed to identify the number of claims for each unique provider zip code in which services for that beneficiary were received. Past methods for defining health service areas have often used counts of utilization events (i.e., hospital discharges) (Baker 2001
), and may, therefore, have been disproportionately influenced by ill patients with high utilization rates. Because historically a relatively small percent of the population uses a high percent of services (e.g., surveys in 1987 and 1996 reveal that the top 1 percent of the population accounted for 27 percent of aggregate expenditures; the top 5 percent of spenders used one-half of all health spending in the same years [Berk and Monheit 2001
]), PCSA definitions could be biased toward the most ill beneficiaries even though primary care serves those who are healthy as well as those with illnesses of diverse resource requirements. Thus, to balance the use rates of the low versus high users, we devised a system of “preference” weighting. Each beneficiary's total use was therefore set at total “vote” of one. The proportion of the beneficiary's total weighted claims located within a particular provider zip code was termed a preference fraction. For example, a beneficiary receiving three services in one provider zip code and two services in another provider zip code would have a preference fraction of 0.6 for the first provider zip code and a preference fraction of 0.4 for the second provider zip code. The sum of that beneficiary's preference fractions, as defined, must equal 1.0. For beneficiaries residing in more than one zip code, their preference vote was apportioned to population zip codes in accordance with the frequency of residence zip codes in their claims.
The patient origin matrix can be viewed as an R×C table with R equal to the number of resident zip codes represented in rows and C equaling the number of provider zip codes represented in columns. The value in each cell of the matrix is a sum of the preference fractions at the intersection of a particular resident zip code and a particular provider zip code. Crude PCSAs were formed by assigning population zip codes to the provider zip codes with the most preference fractions, which indicated the provider zip that received the highest number of vote fractions. The preference of beneficiaries without primary care utilization (35 percent for 1996) was, of course, indeterminate. In some provider zip codes, the beneficiaries received the plurality of their primary care services from a different provider zip code and both provider zip codes were included in the PCSA. This often occurred when a sole provider was adjacent to a zip code with a large medical group.
Population zip codes were reassigned from their primary provider zip code to a provider zip code with a lower total of preference fractions for four reasons. First, in keeping with nearly all previous methods of health service area definition, final PCSAs always constituted contiguous zip codes. To achieve this, 6.4 percent were reassigned. Second, in some instances the beneficiaries residing in the crude PCSA failed to seek the plurality of their primary care (based upon preference fractions) from providers within the same PCSA. This signifies that the method failed to assign the population to the providers they most commonly used. In this instance, 4.4 percent of the zip codes were reassigned. Third, a number of crude PCSAs were identified as having beneficiaries with excessive border crossing, defined as more than 70 percent seeking the plurality of their primary care from outside the PCSA, and their component zip codes were assigned to an alternate PCSA (3.8 percent). Finally, PCSAs with populations of less than one thousand were judged not to have sufficient populations reasonably to support a primary care clinician, necessitating a reassignment of 0.2 percent of the zip codes. Despite these reassignment procedures, 85 percent of population zip codes remained assigned to the provider zip code with the highest preference of the Medicare patients.
In a PCSA ideally defined for measuring primary care resources and utilization rates, the population inside a PCSA obtains all of its primary care from clinicians within the area. Recognizing that there will always be some patients who seek care in other areas, the localization of utilization is a key performance indicator of PCSA definitions and is measured through the preference index. This is the proportion of summed preference fractions for the population residing in a PCSA that occurs in provider zip codes within the same PCSA. Preference indices were calculated from the patient origin matrix after the previously described zip code reassignments. In addition, we calculated mean adjusted preference indices for the PCSAs of each state that were weighted for varying PCSA population size; this is equivalent to the proportion of an entire state's population preferences to a provider within their home PCSA.
Because primary care is thought of as a local medical service, the size of a PCSA is another important characteristic, and one that is positively correlated with the degree of utilization localization. The larger the area, the more localized the primary care activity. Given, however, the local nature of primary care services, smaller areas are more revealing of resources and care patterns. At a certain size, the greater accuracy in the per capita measure of physician supply may obscure heterogeneous primary care availability as areas of locally high and low supply are hidden in the overall rate.
Evaluation with Non-Medicare Utilization
Although it is desirable to define PCSAs using all payer claims data, no such national dataset exists for any purpose. Instead, the project evaluated the generality of the Medicare-defined PCSAs by calculating preference indices with Medicaid and Blue Cross Blue Shield claims in several states. A PCSA with similar Medicare and non-Medicare preference indices is evidence that a non-Medicare population localizes its use in the same service area, an important validation of the PCSA boundaries.
In order to make these comparisons, we used 1995 Standard Medicaid Research Files from HCFA for eight states. The states (Florida, Kansas, Maine, Michigan, Missouri, New Hampshire, Utah, Vermont) were part of the research project's core of so-called pilot states, selected on the basis of interstate geographic differences, of familiarity with the state's health care system, and of state and private agencies willing to work with the project staff. The necessary provider characteristics for claims linkage were, however, available for only five states (Maine, Missouri, New Hampshire, Utah, Vermont). Accurate provider data were not available for Kansas, and the Michigan Medicaid file included widespread errors in the zip codes of enrollee addresses. The Florida ICD-9 diagnosis code data were incomplete. South Carolina was the ninth pilot state and provided Medicaid claims directly from the South Carolina Office of Medicaid, leaving six files for analysis.
Analyses focused on enrollees under age 18 years in order to study a population less selected by health status than others eligible for coverage on the basis of disability. Compared to the Medicare files, the place-of-service and provider specialty fields were relatively incomplete; these fields were used only to exclude claims that were obviously nonprimary care (e.g., pharmacy, specialists). We selected primary care claims by first using the 1995 National Ambulatory Medical Care Survey (National Center for Health Statistics, Hyattsville, MD) to identify the fifty most common primary diagnoses of office-based general pediatricians and family practitioners caring for patients less than 18 years of age. We then used these diagnoses as the Medicaid claims selection criteria for primary care. This definition of primary care utilization was moderately accurate. The diagnoses represented 78.6 percent of the visits nationally provided by these primary care clinicians and, in turn, 61.3 percent of visits with these diagnoses were provided by general pediatricians and family practitioners.
Finally, we evaluated PCSA definitions using a 1996 Michigan Blue Cross Blue Shield (BCBS) file. Michigan BCBS has one of the highest market shares of a commercial carrier in any state (Blue Cross Blue Shield of Michigan 2001
). Claims were selected using the same criteria as Medicare Part B.
In addition to the preference and localization indices, demographic and geographic characteristics were determined using 1999 population data from Claritas, Inc. (San Diego, California). The PCSA geographic size was measured using ArcView 3.2 (ESRI, Inc., Redlands, Washington).