Geography has been a critical component of health care analysis for many decades. In their classic text on health services research, Joseph and Phillips provided a discussion of health care delivery systems around the world, clearly defined the meanings of concepts such as "access" and "utilization," and presented many spatial methods that have been used to analyze them [1]. The text was written when GIS was in its infancy (although developments in computer cartography had been proceeding for several years), but the authors demonstrated how mapping could be used to gain an understanding of the spatial organization of health care providers, such as physicians, and the spatial hierarchy of hospital facilities. They reported on a wide range of methods for measuring accessibility to health services, many of them borrowed from economic geography. These include the location quotient and coefficient of localization, which provide general measures of regional distribution and inequity. Many of the methods used to assess regional accessibility fall into the "distance decay" category, i.e. they involve measurements of cumulative distances between health services and neighbourhoods or other population units and operate under the assumption that distance is a barrier or deterrent to health-seeking behaviour. Other topics covered were health planning and strategies for locating new hospitals and health services.

Many of these methods, such as the computation of physician-patient ratios and location quotients, were demonstrated in Rickets et al. [2]. Rural-urban classification systems were discussed and health care shortage areas mapped. Most of these methods can be applied with geographic information systems.

A geographic information system is an information management system that contains spatially referenced data. Clarke has referred to GIS as 1) a toolbox, 2) an information system, and 3) an approach to science. As a toolbox, a GIS is a software package that contains a variety of tools and functions for processing, mapping and analyzing spatial data. As an information system, it contains a series of databases with observations about features and other entities with known locations. As an "approach to science" it involves the study of the scientific disciplines, such as geography and cartography that have contributed to the development of GIS technology [3].

As an information system, geography thus is the common denominator for disparate data types. Mapping the location of prevention services in relation to HIV incidence, for example, could graphically demonstrate possible gaps in service availability and suggest priorities for locating new service sites. Because maps make complex data more accessible to both experts and non-experts, they can facilitate discussion about issues of access and service needs.

While GIS data are often viewed as maps, the spatial relationships among objects and features in a GIS make it much more powerful than a mere mapping tool. The main reason for this is that spatial data in a GIS are structured in a manner that maintains topological relationships among features such as points, lines and areas. These relationships include adjacency, containment, and connectivity and they allow GIS users to perform analyses among features in a single map layer or multiple map layers [4]. For example, distance could be computed easily between hospitals in one map layer and census blocks in another map layer, to measure accessibility of population to hospitals using a distance decay function. GIS technology can also support spatial query, analysis and modeling functions. Techniques include buffer zone analysis that can estimate the number of persons living within a specified distance of a resource (such as a test site), or the distribution of HIV prevention services in relation to the number of persons living with AIDS.

GIS technology has lead to the enhancement of existing techniques and development of new methods for analyzing health services. GIS can be used to create health service regions based on spatial data using such methods as Thiessen polygons and flow mapping [2]. Spatial interaction models, mathematical programming and GIS network analyses use street data and distance measurements to model flows among patients and health services and are used to allocate patients to services, route emergency vehicles, or strategically locate new facilities [5]. Many of these processes are iterative and can be run many times to examine a host of different scenarios. Of course, GIS analyses would be rather limited without the widespread availability of digital spatial data and health data with geographic identifiers [6].

GIS is increasingly used in public health and health services research. Since 1994, the National Center for Health Statistics has published its bimonthly electronic report, Public Health GIS News and Information. In 1999, the Journal for Public Health Management and Practice devoted two issues entirely to GIS applications. In May, 2003, the Annual Review of Public Health published five articles that, together, constituted a "minisymposium" on the use of GIS in public health. Two issues of the Journal of Medical Systems were devoted to GIS in 2004. GIS has been recognized as an emerging technology in the field of public health informatics [7]. Additionally, new journals, such as the International Journal of Health Geographics have emerged to meet the demand for research in medical informatics and geographic analyses of health issues.

The CDC has made a commitment to utilizing new technologies to improve health information [8]. The potential of GIS has also been recognized by the Department of Health and Human Services (DHHS). The Healthy People 2010 Objective 23-3 is to "increase the proportion of all major national, state, and local health data systems that use geocoding to promote nationwide use of geographic information systems (GIS) at all levels" [9].

Several recent studies have involved the use of GIS in health services research [10-12]. Of renown also are the Dartmouth atlases of health care, which use data from health care claims databases to map and analyze geographic aspects of health care in the U.S. [13,14]. Most geographic studies are more localized, however, such as an analysis of the accessibility of HIV services in Toronto neighbourhoods or GIS-based assessment of physician shortages took place in a 9-county area of Illinois [15,10].

One of the most important potential uses of GIS technology in evaluation and planning of health care services is as a spatial decision support system (SDSS). SDSS provide information to planners and program managers that allow them to make important decisions about resource allocation. They require the development of a spatially-enabled database, a database management system, and a set of analytical tools, such as those found with most GIS software, for solving problems [16]. The first step in the development of a spatial decision-support system is to develop a spatially-enabled database.

The result of our work has been the development of such a database – a dynamic, national spatial database of locations and corresponding geographic service areas of CDC-funded CBOs that provide HIV prevention services. We termed this database the HIV Prevention Services Database. This database is maintained in a GIS and has enormous potential to provide information to program managers for decision making.

Data describing intervention types and persons served, combined with the address of responding CBOs, would by itself yield valuable information about the locations of services being provided with CDC funds for specific populations. However, the intent of this study was to describe service area as well as service location. Geographic service areas can be defined in several ways, each of which has ramifications in terms of data collection issues and analyses:

1. Patient origin. The service area is defined by compiling actual addresses for persons served. Although this approach provides very precise data, it also involves concerns about respondent burden, confidentiality, and data quality. Many HIV prevention programs do not collect address information; consequently, this approach was not feasible.

2. Geographic distance. The service area is defined by the maximum distance from which persons served come to the service. Distance measures are relatively simple in terms of data collection and management. However, because service areas rarely correspond to circular areas described by distance measures, the resulting data can be of relatively poor quality. In some cases, distance measures are converted to administrative units that fall within the specified distance (e.g. all counties that are entirely or partially within a 50-mile radius).

3. Geopolitical boundaries. The service area is defined by naming the states, counties, cities, ZIP codes or other administrative units in which services are provided. These units are familiar to most persons and may already be used by respondents in planning and describing their activities. However, geopolitical units may not correspond to service areas that are defined in terms of neighborhoods, and they are sometimes imprecise, such as when a city boundary spans county lines [18].

Based on discussions among the project team and findings from the pilot test, the study team decided to collect service area data in the form of both geographic distance measures and geopolitical units. We did this because there was no clear precedent as to which method would provide the most useful information, and this would provide us with an opportunity to test both. The questionnaire provided respondents with a cascading set of geopolitical unit responses, from which they could list multiple responses at one or more levels of specificity, i.e., multiple counties or a county with additional cities. This list was, for the most part, geographically hierarchical. If, for example, the respondent checked the box for "entire state" and listed "Missouri," no other geographic area response for Missouri was necessary. Response options for distance included six choices ranging from less than 5 miles to more than 25 miles. Response options for service areas are shown in Table 2.

We defined "service area" as the location of persons actually served. In some of the health services literature, this is referred to as "market area" [19]. This may differ from the target area, for which services were planned. For CDC-funded services, the concept of "service area" provided more useful information. The question was phrased in terms of where persons served live, although we did not ask respondents to consult actual address records when choosing their response. For street and community outreach activities, we instructed respondents to describe the area in which the intervention took place because these activities may be directed at transient populations or persons who congregate in a specific area without necessarily living there.

Our instructions to respondents about service area were to specify it as "the area where the majority (roughly 80%) of people receiving this prevention program live," or, for street and community outreach, "where the majority of activities took place." This wording was intended to avoid responses that were skewed toward large service areas by a small number of service users or activities outside the usual service area. In the pilot test, the study team found that this wording elicited responses that more closely represented actual activities.

Each survey package included a custom-made one-page color reference map created for that CBO and generated by an automated GIS routine. The map showed two views of the area surrounding the CBO's location: one identifying cities, counties, and major roads within a 30-mile radius; the other showing a more detailed view of ZIP codes and towns within a 5-mile radius. In both views, we plotted concentric circles at set distances to provide a spatial frame of reference. We based our decision to include these maps on the San Diego pilot test, in which respondents completed service area questions twice: first without a reference map, and then with it. Using a reference map improved data quality in several ways:

1. Completeness. Respondents named more cities served when looking at a map that included names of all cities in the county.

2. Accuracy. Estimates of distance from the CBO location were more accurate when respondents consulted a map showing distance in 5-mile increments.

3. Precision. Respondents described service areas in terms of specific ZIP codes within the city rather than the entire city when using a map showing ZIP code boundaries.

We converted the text files from the data entry process to a series of 10 Microsoft Access 2000 tables. These 10 tables were developed to normalize the data, (i.e., group them into tables in a formalized procedure to eliminate duplication of information and provide flexibility in table structure for future additions or changes) and to allow linkage to GIS map files via GIS software. Full details of the database design are described in a separate report to CDC and are beyond the scope of this paper [20]. We discuss three important tables in the database, however. These are shown in the Figure 3 schematic, which uses a fictitious CBO. Field names have been changed for readability.

The first table, CBO, contains a master list of CBOs and includes the following information: CBO identifier; name, address and contact person; and information about survey responses. This information was used for survey administration and geocoding.

The FIPS codes for all geographic service area entities (i.e. state, county, city/town, ZIP code and/or Indian Reservation) were stored in a second table, GEOGAREA. This table contained four fields: 1) the CBO identifier, 2) the program identifier (many CBOs had multiple programs), 3) the FIPS type (e.g. state, county, city, ZIP, tribal lands) and 4) the actual FIPS codes. Each geographic unit that represented a portion or all of a service area for a particular program was stored as a single record. In the Figure 3 example, the Fictitious CBO was assigned an identifier of 15015. The service area of this CBO's Program #1 covered three zip codes. Data in the FIPS_TYPE field indicate which GIS base map (i.e., state, county, city, ZIP code, or reservation) to link to. For example, a FIPS_TYPE of 4 indicates that the linkage is to the national ZIP code map layer. The values in the FIPS_CODE field are actual ZIP codes, which can be queried and displayed with the GIS.

A third important table, PROGRAM contains all of the non-geographic information for each program, i.e. intervention type, risk populations, race/ethnicity and funding source. This table is linked, via a combination CBO/program identifier, to the geographic tables. This table also stored values provided for the last survey question, in which respondents were asked to indicate the distance within which the majority of people served lived.

We used a suite of Environmental Systems Research Institute, Inc. (ESRI, Redlands, CA) GIS software products for all spatial data processing and analyses, including ArcGIS 8.12, ArcMap and ArcCatalog. However, the final product of this research – a dynamic spatially-enabled database to be used by CDC program managers – was set up for use in ArcView 8.

We integrated the survey data in the Access database with a series of GIS map layers for subsequent mapping and analysis. These included U.S. states and territories, counties, cities and towns, American Indian reservations, and ZIP code area boundaries. These GIS map layers were derived from two sources: 1) generalized U.S. Census Topologically Integrated Geographic Encoding and Referencing (TIGER) 2000 Arc/Info export files, obtained from the U.S. Census web site; and 2) the ESRI Data and Map series, Version 8.1, which came bundled with ESRI software.

We made some enhancements to the original map layers to incorporate all survey responses. The ZIP code area boundary layer includes some small buffer polygons of ZIP code points that were added for this project. A number of the cities and towns that were identified by the survey participants did not exist in the city/town map layer, so we augmented the places map layer with places found in the online U.S. Geological Survey (USGS) Geographic Names Information System(GNIS). Lastly, a special areas layer was created manually from other background data sets for a few areas specified by survey participants that did not match any of the other background layers.

All responses about geographic services areas were matched to one or more of the geographic boundary files described above. A different procedure was used to develop map layers of CBO and program locations. The CBO and PROGRAM tables in the Access database contain addresses for CBOs and their programs. These addresses were used to derive the CBO and program point locations in geographic coordinates (i.e., latitude and longitude), such as those displayed in Figure 1. CBO and program addresses were address matched by a vendor. Response codes were linked to the address matched CBO data, so response status (i.e., whether the CBO responded to survey or not) of each CBO could be queried and mapped.

The GEOGAREA table contains information about all geographic entities that were indicated, by respondents, to be part of a geographic service area. Responses about geographic distance were stored in the PROGRAM table and linked to the map layer of program locations. The ArcView Buffer Wizard was used to buffer each program point by the corresponding distance estimate to create a new map layer showing service areas based on distance.

The primary result of this project is the HIV Prevention Services Database, a dynamic, spatially-enabled database that provides CDC with a wealth of information about HIV prevention services that it funds, and a large potential for geographic modeling, analyses, and mapping. This database handles overlapping geographies, risk populations and prevention services. In order to make it user-friendly for CDC program managers, we provided CDC with an ArcView (.mxd) application that automatically loads all of the spatial data (i.e. shapefiles) and tables needed for analysis and mapping. Relationships among tables (i.e. "joins") needed for query and analysis are also maintained in this ArcView application. Due to the wide range of potential database queries, we developed a Visual Basic for Applications (VBA) query tool that makes it easy for users to structure a query based on intervention type, race/ethnicity and risk population. With this tool, the user has the option of mapping CBO and program locations, plus their geographic service areas. The query tool interface is shown in Figure 4.

The HIV Prevention Services Database presents a wide range of query and display capabilities, based on responses to the survey. For example, Figure 5 shows the geographic service areas of all programs that provide interventions to Hispanic/Latino populations. While the map is national in scale, the zoom and query functions in a GIS allow users to examine geographic areas at any scale. This query was based on program responses to questions about intervention type and populations served.

Queries can be based on geography as well. The question, "Which HIV prevention services are being provided by CDC-funded CBOs in the state of Rhode Island" would produce a map of Rhode Island CBO locations and their service areas and a wide range of information about types of service and risk populations in connected database tables.

The HIV Prevention Services Database is being used by CDC researchers. One analysis has focused on the geographic distribution of services at the national level and another on services to specific populations, such as African Americans.

Many CBOs indicated that they provided prevention services to an entire state. In many cases, this did not seem feasible, and concerns were raised about the integrity of these responses. We developed a set of procedures for confirming the validity of state responses that included 1) using the distance values in the last survey question for validation; 2) a consideration of the size of the state (statewide coverage in Rhode Island is more feasible than that of Texas, for example); 3) examining the type of intervention (e.g. prevention case management vs. health communications;) and 4) telephone callbacks to CBO program administrators by CDC staff.

In some cases, geographic entities provided by survey respondents simply could not be located in a geospatial database or even an atlas or gazetteer. The most common of these were the ZIP codes. Some CBOs provided ZIP codes that did not exist in the United States Post Office database. Thus, for some CBOs, service area data are missing or incomplete.

Some of the ZIP codes identified by survey respondents did not exist in the ZIP code polygon (area) GIS map layer, but did exist in another GIS layer of points only (i.e., represented by a singlelatitude/longitude coordinate). We made the assumption that these "point only" ZIP codes represented very small ZIP code areas. These ZIP codes were given "area" coverage through the creation of 0.1-mile buffers around their representative points.

We used a series of GIS queries and logical consistency checks to identify data anomalies. Each time an inconsistency was noted, we examined the original surveys. In a handful of cases, the coders had misinterpreted the respondent's handwriting, and we made corrections.

In spite of some of the challenges we encountered, we successfully developed methods to obtain primary information about CDC-funded HIV prevention services in the U.S. and its territories and were able to develop a dynamic GIS database of CBO locations, service areas and prevention services that is being used by CDC staff to perform analysis and make program decisions. This database, the HIV Prevention Services Database, was delivered to CDC in May 2002.

We need to caution, however, that the response rate to the survey was 70%. While this is a high response rate, we realize that program and service area data are missing for 430 CBOs. Any comprehensive analysis of service provision must take into account the locations of these non-responding CBOs. Additionally, CBO-provided services funded by CDC by no means make up the total of HIV prevention services in the U.S.

The CDC HIV Prevention Services Database was developed to enable CDC researchers to plan and evaluate CBO-provided HIV prevention services. While CDC is using this database primarily to identify gaps and overlaps in service, its potential is broad and includes the following applications:

• geographic analyses of HIV-prevention services to racial and ethnic minorities,

• an examination of HIV prevention services in the context of health disparities as a follow-up to work by Krieger et al. [21],

• use by local health agencies to determine how to integrate CDC-funded services with other community prevention services,

• regional studies of HIV prevention services for areas such as Appalachian Regional Commission counties, and

• analysis of CDC-funding levels vs. assessment of need, based on HIV/AIDS rates and risk populations.