Place, Space, and Health: GIS and Epidemiology

Place. Area. Neighborhood. Latitude. Longitude. Distance. These geographic terms are increasingly finding their way into the epidemiologic literature, as advances in geographic information system (GIS) technology make it ever easier to connect spatially referenced physical and social phenomena to population patterns of health, disease, and well-being. 1–3 Indeed, links between location and health have long captured the imagination of perceptive observers. Consider the Hippocratic treatise, “Airs, Waters, and Places, ” written about 2, 400 years ago, which roundly (and rather deterministically) declared: “You will find, as a general rule, that the constitutions and habits of a people follows the nature of the land where they live. ”4, p. 168 Early 19th century research decisive to epidemiology’s development as a discipline 5 likewise looked to geography to discern etiologic clues. For example, neighborhood mortality rates were linked to poverty rates, 6, 7 and the risk of outbreaks of yellow fever was analyzed by distance from docks. 8, 9 A celebrated late 19th century text went so far as to call epidemiology the science of “geographical and historical pathology, ”10, p. 2 a definition embraced by Wade Hampton Frost 11, p. 494 and other prominent epidemiologists well into the early 20th century. Yet, despite epidemiology’s longstanding concern with “time, place, and person”12 (or, perhaps more accurately, “time, place, and population”13), “place” had receded into the background by the mid-20th century, conceptually unmoored from increasingly influential etiologic frameworks based on characteristics of the individual. 5, 13 Fortunately, GIS has contributed in recent years to a reviving awareness that any epidemiologic explanation worth its salt must encompass geographic—and temporal—variations in population health. Discussions are being enlivened by new research drawing on multilevel frameworks and methods exploring the public health salience of “place. ”5, 13–18 GIS unquestionably offers epidemiology a wonderful new tool. If poorly used (or poorly made), however, a well-intended tool can do more damage than good. Underscoring this concern are two issues raised by four papers in this issue on aspects of GIS. 19–22 First, “completeness” in geocoding does not equal “success”. Accuracy—and choice of geographic level–matters as much if not more. Research from our Public Health Disparities Geocoding Project, 23 for example, has found not only significant variability in geocoding accuracy by diverse commercial firms 24 but also introduction of major bias by spatiotemporal mismatches between census-defined areas and zip codes. 25 Extending this work, Hurley et al 19 powerfully demonstrate that serious misclassification can occur if post office boxes are geocoded to their zip code centroid and then analyzed as if the centroid were where people actually live. Their study also found that persons with post office boxes differed from those with residential addresses on such key characteristics as age, race/ethnicity, and whether data on tumor stage at diagnosis were missing. McElroy and colleagues spell out the costs of attempting to geocode every single address, using multiple methods of unknown accuracy (a time-consuming strategy that cost 12, 500 to geocode approximately 15, 000 records 20). A likely preferable, albeit less complete, alternative would be to use a single cost-effective method with verified high accuracy (eg, one costing approximately 550 for 15, 000 addresses 24), and then thoughtfully consider how selection bias could potentially affect results. “If poorly used (or poorly made), a well-intentioned tool can do more damage than good. ” Second, requirements for geocoding accuracy and types of spatially linked data will vary depending on study needs. Investigations such as Floret’s 21 on cancer incidence and incinerator emissions require precise distance between a given address and a specified location. Such studies can benefit from methodologic research like Bonner’s 22 on positional accuracy, which reassuringly found generally good agreement between latitudes and longitudes assigned by a widely used software package and by a global positioning system receiver. Positional accuracy, however, is only one piece of the picture. Studies investigating links between an area’s characteristics and health additionally face the challenge of defining relevant areas, choosing apt area-based measures, and delimiting appropriate exposure periods. 15–18, 23–26 These choices depend on the study’s objective, eg, enhancing public health surveillance, or delineating and testing particular etiologic pathways. 23 In summary, and as usual, both methodological and conceptual precision matter, as does attention to practical details of cost and time. There are no ready-made answers. If GIS is to generate valid data for testing hypotheses about population health, epidemiologists will need to document the validity of our GIS methods and provide conceptual justification for the geographic levels we choose to study, as well as for the measures we employ. ABOUT THE AUTHOR Nancy Krieger is a social epidemiologist and Associate Professor in the Department of Health and Social Behavior at the Harvard School of Public Health. She has been working with geocoding and census data since 1985 to document and analyze health disparities involving race/ethnicity, class and gender. Current projects include using GIS to improve monitoring of United States socioeconomic inequalities in health.