Using Geographic Information Systems (GIS) to Understand a Community's Primary Care Needs

The Practice-Based Research Network

To develop our model we used the resources of the Mecklenburg Area Partnership for Primary Care Research. This regional PBRN is based at Carolinas Health care System (CHS), the third largest vertically integrated hospital system in the United States. The CHS system provides the majority of care for our underserved community. Based on patient volume and community-wide data we estimate that this system provides more than 85% of emergency care and 80% of primary care services to the community's uninsured patient population. The primary care safetynet for this project was defined using 4 of the system's primary care clinics that provide the majority of services for the community's uninsured patients.

The methods described were developed by Mecklenburg Area Partnership for Primary Care Research as a way of carefully targeting interventions designed to improve access to primary care, with the goal of ameliorating disparities and improving the health of the community. The research network brought together ambulatory providers, community members, and research team members as a community advisory board (CAB) using principles of community-based participatory research.⁷ The network's CAB consisted of the following members: The Carolinas Medical Center (CMC) Ambulatory Clinic System (CMC Eastland Family Practice Clinic, CMC NorthPark Clinic, and CMC Biddle Point Clinic); Our Lady of Guadalupe Church; The Bethesda Clinic (a local free clinic); The Mecklenburg County Health Department; Department of Family Medicine; University of North Carolina at Charlotte Geography and Earth Sciences Department; and the Latin American Coalition. Advisory board involvement spanned from development of the research question to selection of attributes to oversight of research processes and review of results. Thus, the process involved both partnering with and tailoring to the needs of the community.

Analytic Hierarchy Process

The AHP is a structured technique for making complex decisions.^14,15 Rather than giving a “correct” decision, the AHP helps decision makers find the solution that best suits their perception of the research problem. The research question is translated into a hierarchy of more easily comprehended subproblems, each of which can be analyzed independently. Once the hierarchy is built, the decision team systematically evaluates its various elements by comparing them to one another. When making comparisons the decision makers can use data or personal judgments about each elements’ relative importance. These evaluations are converted to numerical values that can then be compared.

The decision makers can assign weights to each factor so they can be combined quantitatively on a map. For example, if we have 5 factors of equal weighting, each factor would be multiplied by 0.2. Then the factors would be added together to give a final score for each census tract. If the weightings were different for each factor, this would enable the CAB to emphasize key factors with greater weightings in the overall map.

For this project, the available community- and patient-level data were reviewed to identify what information was most pertinent to answer a critical question in the process: What are the best indicators of primary care need for our community? The research team and advisory board used an iterative process of reviewing the available data, selecting attributes that were deemed important, and reviewing maps of each data element before selecting 5 key attributes. Identifying these attributes was a key first step in answering the research question of what geographic areas within a defined community would receive the greatest benefit by increasing access to primary care services.

Each variable was reviewed by the team for its ability to describe the community's primary care needs and contribute to the answering of the research question. This iterative process led to multiple assessments and weightings of different variables, including (1) the selection of population density; (2) the identification of uninsured/Medicaid insurance types as risk factors for primary care need; and (3) the choice of percentage of population using a primary care safety-net clinic as a protective variable instead of using the primary care service area alone.

The 5 selected attributes were standardized to values between 0 to 1, and multiple sets of composite maps were created based on several different weighting criteria. After evaluating the different weighting schemes, the team chose equal weighting for all variables, with utilization of the primary care safety-net being given an inverse but equal weight as a protective factor.

Multiple Attribute Assessment and Evaluation

MAAE methods analyze multiple attributes and combine data elements to positively influence decision quality.¹⁶ For this project we used MAAE methods to weigh multiple attributes selected through the AHP process (described above) to gain a better understanding of community-wide primary care needs. The members of the research team and CAB ranked each of the selected attributes according to its perceived importance. The attributes were rated on a scale (0 to 1), with 1 being the highest level of importance. The attributes were then combined and maps showing the data were created and reviewed. After reviewing multiple sets of weights for each attribute, the group selected a system of equal weighting. Discussion among the research team and CAB concluded that no single factor deserved differential weighting. As such, all indicators were considered equally reflective of primary care need in the community.

Geographic Information Systems (GIS)

After institutional review board approval for the study was obtained, data for all ED and primary care clinic visits in 2007 was extracted from the shared CHS databases. The initial dataset contained just more than 367,000 records. Data for patients living outside of the county or lacking an address were excluded, which resulted in approximately 187,000 ED visits and 50,000 primary care patients being included in the analysis. Census-based demographic data for Mecklenburg County projected for 2007 was obtained from Claritas (San Diego, CA). All patient address data were geocoded using ArcGIS (ESRI Products, Redlands, CA) greater than a 97% match rate for both clinic and ED data.

To ensure patient privacy, the original addresses and any other information that could identify the patients (other than the geocode) were stripped from the data set before further analysis. In addition, all data were reviewed to identify any census tracts containing fewer than 50 individuals from any one racial or ethnic category. When identified, these data points were either removed or averaged into an adjacent census tract. All maps created from the dataset were at the level of the census tract or higher, and no dot maps were created for this project.

New York University Emergency Department Algorithm

The New York University (NYU) Algorithm was developed by the NYU Center for Health and Public Research to classify ED visits and identify inappropriate utilization of the ED based on the patients’ diagnosis.¹⁷ The categories include (1) nonemergent; (2) emergent but treatable through primary care; (3) emergent with ED care needed but preventable or avoidable with access to primary care; (4) emergent with ED care needed and not otherwise preventable; and (5) other categories, including injury, mental health, drug/alcohol related, and unclassifiable. We applied the NYU ED algorithm (version 2.1 for Microsoft Access) applied it to our geocoded ED data. To facilitate this process, our ED data were reduced to only 2 fields (the geocode and International Classification of Diseases version 9 code) before application of the algorithm. The resultant data set provided a geocode linked with a probability (from 0 to 1) for each of the categories described above. For this study we combined the probabilities for categories 1 to 3 because ED visits in each of these groups were considered treatable through primary care or preventable with appropriate access to primary care. These combined categories for issues treatable through primary care or preventable ED visits are also called “ambulatory care sensitive” (ACS).

MAPCATS

A flow diagram for the MAPCATS process can be seen in Figure 1. After the research question was defined, the descriptive community data—including population density, household income, and race/ethnicity—were imported into ArcGIS (ESRI Products), allowing the team to assign weights to each attribute before creating maps. Then the AHP process was used to select and refine attribute selection from the available data fields (see Figure 2). Members of the research team and the CAB reviewed maps of individual attributes to confirm that the data fit their perceptions of the community and that these factors would be valid indicators of primary care needs. Next, combined maps were assembled based on the MAAE process (described above) to determine the appropriate weightings for a final map. Finally, combination maps incorporating the weightings determined by the group process were assembled to show the primary care needs of the community.

• Download figure
• Open in new tab

Figure 1.

Overview of the Multiattribute Primary Care Targeting Strategy (MAPCATS). This process started with engagement of the community advisory board that guided and oversaw each step of the research process. The team selected the research question of “What geographic areas are in need of increased access to primary care services?”

• Download figure
• Open in new tab

Figure 2.

Overview of the data processing for the Multiattribute Primary Care Targeting Strategy (MAPCATS).

To validate the results of the model, the individual variables selected were combined using factor analysis.¹¹ The final component scores (the result of one or more components combined) were used to create an overall combination map to compare with the final target map.

Map Preparation for Each Attribute

Socioeconomic data were represented by median household income—an easily obtained community indictor available from the census-based Claritas dataset. These data were mapped by census tract, as seen in Figure 3A. Income was considered an important indicator of primary care need, and targeting low-income community members would provide a means of helping to overcome health disparities and improving community health.^19,20

• Download figure
• Open in new tab

Figure 3.

Maps showing each attribute used to assess community-wide primary care need. A: Median household income by census tract. B: Population density by census tract. C: Emergency department patients, uninsured or using Medicaid, by census tract. D: Patient emergency department utilization for illness preventable or treatable through primary care, by census tract. E: Patient utilization of safety-net clinics by census tract.

Population density was identified by dividing the total population in each census tract by the census tract area in square miles. Data were then mapped using ArcGIS (ESRI Products) to create Figure 3B. Population density was considered an important attribute because interventions to improve health care access need to be in areas that can impact the largest number of at-risk community members. Population density maps were checked against the hospital's patient data, and rates of patient ED utilization were found to be directly proportional to the overall population density.

Insurance status was drawn from the CHS clinical databases. This data field was sorted based on input from the research team members and primary care providers into 5 categories: uninsured, commercial, Medicaid, Medicare, and other. A separate analysis, comparing each insurance type and ED utilization for ambulatory care-sensitive diagnoses, revealed that lack of insurance and Medicaid insurance coverage were significantly linked to inappropriate ED utilization (unpublished data). From this analysis, the number of patients within either of these 2 insurance categories were combined and divided by the total number of patients for each census tract. This data were mapped using ArcGIS (ESRI Products) to produce Figure 3C.

ED utilization patterns were determined using the NYU ED algorithm to identify ACS ED visits. Our assumption was that these ACS visits could have been treated in a primary care setting or avoided if patients had access to primary care in a more timely fashion. The ACS scores provided a probability between 0 to 1 for each patient visit. These scores were combined at the census-tract level and mapped using ArcGIS (ESRI Products) to provide Figure 3D. To confirm the validity of this data, maps were also created for non-ACS visits. The maps showed patterns that were the direct inverse of the ACS ED maps incorporated into this model.

The final attribute used in the MAPCATS process was patient utilization of the community's primary care safety net. Data from the 4 primary care clinics that provide the majority of care for our community's uninsured patients were analyzed. The percentage of the population using the safety-net clinics was determined by dividing the number of patients in each census tract by the total population of that tract. This information was then mapped and can be seen in Figure 3E. This attribute was added to the model as a protective factor for primary care need, so the areas with a higher percentage of the population accessing services were considered to be less likely to need additional primary care access.

Creation of Composite Map

The numeric values for all the variables in each census tract were added together as a measure for the tract's perceived need for additional primary care services. The cumulative total was then obtained, and the census tracts with values greater than 1 SD above the mean were selected as target areas that would probably benefit the most from additional primary care capacity (Figure 4). The areas that scored highest for primary care need were located to the east of the city center and were near one of the ambulatory care clinics. Further analysis revealed that this clinic's service area did not include many of the census tracts within its geographically located neighborhood as would be predicted by this model. This finding led to a secondary project that sought to determine why this mismatch occurred and how it could be rectified.

• Download figure
• Open in new tab

Figure 4.

Composite map showing overall community primary care need. A single map shows the final output based on combining the 5 selected attributes describing primary care need in Figure 3.

The 5 individual variables were combined using factor analysis¹¹ to validate the results presented. The analysis resulted in the creation of 2 components that explained 85% of the variance of the variables. The socioeconomic component combined ED utilization, safety-net clinic utilization, percent of uninsured/Medicaid patients, and median household income (Table 1). The density component consisted only of population density. The 2 components were combined together based on the number of variables contributed to each component. The socioeconomic component contained 80% of the variables and the density component contained 20% of the variables. The combined factor scores were then standardized. The results 1 SD above the mean are illustrated in Figure 5 and target more than 85% of the same census tracts, as indicated in Figure 4.

• Download figure
• Open in new tab

Figure 5.

Factor analysis composite map showing overall community primary care need.