Geographic Variation of Pediatric Burn Injuries ... - Wiley Online Library

ACAD EMERG MED

d

July 2003, Vol. 10, No. 7

d

743

www.aemj.org

CLINICAL PRACTICE Geographic Variation of Pediatric Burn Injuries in a Metropolitan Area Kristine G. Williams, MD, MPH, Mario Schootman, PhD, Kimberly S. Quayle, MD, Jim Struthers, BA, David M. Jaffe, MD Abstract Objectives: To use a geographic information system (GIS) and spatial statistics to describe the geographic variation of burn injuries in children 0–14 years of age in a major metropolitan area. Methods: The authors reviewed patient records for burn injuries treated during 1995 at the two children’s hospitals in St. Louis. Patient addresses were matched to block groups using a GIS, and block group burn injury rates were calculated. Mapping software and Bayesian analysis were used to create maps of burn injury rates and risks in the city of St. Louis. Results: Three hundred eleven children from the city of St. Louis were treated for burn injuries in 1995. The authors identified an area of high incidence for burn injuries in North St. Louis. The filtered rate contour was 6 per 1,000 children at risk, with block

group rates within the area of 0 to 58.8 per 1,000 children at risk. Hierarchical Bayesian analysis of North St. Louis burn data revealed a relative risk range of 0.8771 to 1.182 for census tracts within North St. Louis, suggesting that there may be pockets of high risk within an already identified high-risk area. Conclusions: This study shows the utility of geographic mapping in providing information about injury patterns within a defined area. The combination of mapping injury rates and spatial statistical analysis provides a detailed level of injury surveillance, allowing for identification of small geographic areas with elevated rates of specific injuries. Key words: geographic information systems; spatial statistics; pediatric injuries. ACADEMIC EMERGENCY MEDICINE 2003; 10:743–752.

Childhood injury is a local and national public health issue. Between 1995 and 1998, unintentional injuries were the leading cause of death for children between the ages of 1 and 14 years in the United States.1 Effective injury prevention efforts begin with the study of injuries. A variety of methods and data sources have been used as surveillance tools: national health statistics; hospital-based trauma registries; statewide hospital discharge data; police accident reports; and fire department, emergency medical services (EMS), and medical examiner records.2–8 Such surveillance systems provide general geographic or demographic information upon which to base prevention efforts. Traditionally, investigators have targeted interventions to cities, communities, or speci-

fic populations, such as schoolchildren or parents of children in pediatric practices, with varying levels of success.9–13 Identification of discrete geographic areas where injury incidence is elevated may allow for more focused prevention efforts. A recent study of burn injuries used this approach. Based on medical examiner records and medical record reviews of all deaths from residential fires in Oklahoma City, investigators identified a 24square-mile area with the highest rate of burns related to residential fires. Following the implementation of a smoke-alarm-giveaway program targeted to this area, fire-injury rates showed a significant decline.8 The success demonstrated in this study suggests that interventions aimed at discrete geographic areas with an elevated incidence of certain types of injuries might also have an important effect in reducing morbidity and mortality from childhood injuries. Generating maps to study disease occurrence has long been a part of descriptive epidemiology.14 The goal of mapping is to depict the geographic distribution of disease or injury rates, and to identify areas with high or low rates. This can be done through basic mapping strategies in which incidence is plotted for a geographic area and rates are calculated based on population, as demonstrated in the previously mentioned study by Mallonee et al.8 Similar mapping techniques also have been used successfully to study pediatric pedestrian injury.5,6 A major limitation of this type of mapping method is that in small areas, rates can fluctuate over time because of low incidence of

From the Division of Pediatric Emergency Medicine, Department of Pediatrics (KGW, KSQ, DMJ), the Division of Health Behavior Research, Departments of Pediatrics and Medicine (MS, JS), Washington University School of Medicine, St. Louis, MO. Received March 28, 2002; revision received December 18, 2002; accepted December 23, 2002. Presented in part at the Ambulatory Pediatric Society/Society for Pediatric Research Conference annual meeting, Boston, MA, May 2000; and at the SAEM annual meeting, San Francisco, CA, May 2000. Supported by the University of Iowa Injury Prevention Research Center, which provided a pilot study grant. Address for correspondence and reprints: Kristine G. Williams, MD, MPH, Division of Pediatric Emergency Medicine, Box 8116, St. Louis Children’s Hospital, One Children’s Place, St. Louis, MO 63110. Fax: 314-454-4345; e-mail: [email protected]. A related commentary appears on page 780.

744 disease or injury or because of changes in population density. Because of the relative infrequency with which certain injuries, such as burns, occur among children, alternative statistical methods are needed to stabilize the rates and reduce the possibility of drawing conclusions based on elevated rates with few injuries. Use of a geographic information system (GIS) in combination with spatial statistical analysis is a way to address such limitations. A GIS is a computer information system that works with data referenced by spatial or geographic coordinates, allowing for storage, analysis, and manipulation of data. By creating filtered injury rate contours through computer analysis of geographic injury data, a GIS is able to calculate injury rates that are less influenced by small numbers of injuries, since the methods are not bound by traditional geographic boundaries. By taking into account the rates in adjacent areas, injury rates can be estimated for a geographical area that may be larger than the area the actual data represents. In this study, we used a data set of pediatric burn injuries to demonstrate the ability of a GIS to identify discrete geographic areas at high risk for this type of injury. We chose to analyze burns because previous work showed a high incidence of this type of injury in Missouri.15 We then used spatial statistics to analyze the risk of injury based on the geographic analysis. Knowing with increased precision where injuries

Williams et al.

d

GEOGRAPHIC VARIATION OF PEDIATRIC BURNS

occur may allow for future study of geographic or demographic risk factors associated with injury.

METHODS Study Design. This was a descriptive study based on retrospective analysis of burn injuries in St. Louis, Missouri, during 1995. A database of all pediatric burn injuries treated at the study hospitals was created from review of medical records. Approval for this study was obtained from the Washington University and St. Louis University Medical Center Institutional Review Boards. Study Setting and Population. The study sample consisted of children evaluated for a burn injury in 1995 at St. Louis Children’s Hospital (SLCH) or Cardinal Glennon Children’s Hospital (CGH). SLCH and CGH are the two tertiary care pediatric centers for St. Louis and the surrounding areas. Although there are several community hospitals that provide care for some pediatric patients, these two hospitals serve the majority of pediatric patients in the area. Three community hospitals, Christian Northeast, Christian Northwest, and Missouri Baptist, had no pediatric burn injuries reported in 1995. Figure 1 shows St. Louis City, SLCH, CGH, and other area hospitals. Children between the ages of 0 and 14 years evaluated in the emergency department or admitted

Figure 1. Map of St. Louis with the study hospitals: A ¼ St. Louis Children’s Hospital; B ¼ Cardinal Glennon Children’s Hospital.

ACAD EMERG MED

d

July 2003, Vol. 10, No. 7

d

745

www.aemj.org

for a burn injury at SLCH or CGH in 1995, and who lived the City of St. Louis at the time of their injury, were included in the study. A total of 595 patients were evaluated for burn injuries in 1995 in the two study hospitals. Seventy-six patients lived outside of Missouri and were excluded from further analysis. Of the remaining 519 patients, 201 lived outside the City of St. Louis, and for seven patients the street address did not allow geocodes to be assigned due to incomplete medical record information. A total of 311 patients were included in the final analysis for the City of St. Louis (Table 1).

Data Analysis

Study Protocol. We first conducted a computerized search of International Classification of Diseases (ICD9) diagnosis codes to identify medical records for review. Study investigators reviewed individual records and collected information for each patient, including age, race, gender, zip code, street address, ICD-9 diagnosis code, and external cause of injury code (E-code). Socioeconomic data for the study area were taken from the 1990 U.S. Census, and included total population of the census tracts, 0–14-year-old population in divided age groups, percent non-white population, percent unemployment, percent of persons living below the poverty level, percent of children in poverty, median income, median grade level for persons over 25 years of age, and median housing age. This information was used to identify the total at-risk population and to compare study sample demographics with baseline demographics for the city of St. Louis. In order to determine whether the study data were representative of injury data reported to the Missouri State database, we compared rates of burn injuries. Since state data are reported by zip code, burn injuries identified from retrospective chart review were separated into zip codes. The percentage capture rate for each zip code was calculated from the ratio of the study data to the Missouri State data (Figure 2).

Following the work of Rushton and colleagues,16,17 a grid was constructed to cover the city of St. Louis using the block-group centroids as the grid locations. There were 988 grid locations (block-groups) in the study area. From each grid location, a circle of 0.75mile radius was drawn and the numbers of burns (numerator) and children 0–14 years of age (denominator) were calculated within this radius. Thus, an observed burn injury rate was calculated at each grid location. A continuous spatial distribution of burn injury rates was calculated by interpolating the observed burn rates at the grid locations. The interpolated burn injury rates (contours) may cross between two grid

Burns in the City of St. Louis. Burn injuries were address-matched and assigned to block-groups by geocoding subject addresses using Atlas GIS (v4.0, ESRI, Redlands, CA) and the Etak (now Tele Atlas, Tele Atlas North America, Menlo Park, CA) Eagle Geocoder, available on the Internet. Block-group burn injury rates in the city of St. Louis vicinity (latitudes 38.526386 to 38.787223, longitudes –90.361976 to –90.180104) were calculated from the block-group subject counts and the 1990 Census Bureau tabulations for children 0–14 years of age.

TABLE 1. Characteristics of the Study Population of Missouri, St. Louis, and North St. Louis Missouri (n ¼ 519)

St. Louis (n ¼ 311)

North St. Louis (n ¼ 196)

Gender Male Female

280 (53.9%) 239 (46.5%)

170 (54.7%) 141 (45.3%)

Race African American White Other

375 (72.2%) 136 (26.2%) 8 (1.5%)

265 (85.2%) 39 (12.5%) 8 (2.6%)

192 (98%) 3 (1.5%) 1 (0.5%)

Age 0–4 years 5–9 years 10–14 years

351 (67.6%) 98 (18.9%) 70 (13.5%)

202 (65.0%) 61 (19.6%) 48 (15.4%)

118 (60.2%) 41 (20.9%) 37 (18.9%)

99 (50.5%) 97 (49.5%)

Figure 2. 1995 burns reported by the study hospitals as a percent of state data reported by zip code.

746 locations. The result is a map with isolines—shown as lines crossing the block group rates—that displays interpolated values for burn injury rates (Figure 3). The rates are the same at the respective isolines and are expressed as the number of burns per 1,000 children 0–14 years of age. We used the Distance Mapping and Analysis Program software (University of Iowa, Iowa City, IA) to create the grid for the study area and to calculate the interpolated rates.18 ArcView (version 3.2a, ESRI) was used to construct all maps. Burns in North St. Louis. We focused further analysis on North St. Louis, defined as the area of the city north of Highway I-64, because we found this to be an area where the burn incidence was elevated, and because we captured the majority of burn injuries reported to the state for this area, approximately 86% (Figures 2 and 3). The burn rate at the census tract

Williams et al.

d


level was used to describe the small area variation in North St. Louis. Because the rates were frequently based on only a few burns and the variation in the rates often exceeded that of the routinely used statistical distribution, we employed Bayesian analyses. Bayesian analyses are starting to become part of mainstream statistics.19 The most important difference between the traditional frequentist approach and Bayesian analysis is that the latter method incorporates prior knowledge (distribution) about the issue at hand in conjunction with the observed data, in order to calculate a posterior distribution. In contrast, the frequentist approach relies only on the data collected.19 In the case of mapping burn injuries, we considered prior information about the variability of the rates in North St. Louis in addition to the observed rates in this area. Additionally, Bayesian methods incorporate the influence of the burn rates of

Figure 3. Burn rate contours. *This area represents the filtered rate contour of 6 burn injuries per 1,000 children aged 0–14 years at the line border. Although there are other regions depicted with similarly high block-group rates, this is the largest area and is the only one found where there is a relatively high population density (see Figure 4).

ACAD EMERG MED

d

July 2003, Vol. 10, No. 7

d

747

www.aemj.org

neighboring census tracts and the posterior estimate is adjusted accordingly. We constructed a Bayesian model, which included the area sociodemographic covariates. For this Bayesian model, the geographical variation may either be completely spatially unstructured (heterogeneity), spatially structured (clustered), or both. A spatially structured model indicates that geographically close census tracts tend to have similar rates, while a spatially unstructured model indicates that the rates vary independently of the rates in neighboring census tracts. Because it is often difficult to choose between models with just a heterogeneity or clustering component, we have included both terms in the model. Thus, the model is given by the sum of the ðiÞ overall mean m, the heterogeneity random effect bH ðiÞ and the clustering random effect bC , respectively, with variances s2H and s2C .20 All these terms are assumed a priori to be independent. Formally, the ðiÞ ðiÞ general model is given by: log(uðiÞ ) ¼ m þ bH þ bC þ bxðiÞ where uðiÞ is census tract i specific relative risk (RRFB), which is calculated from the ratio of the observed and expected number of burns, and bxðiÞ is the effect of the sociodemographic covariates. The expected number of burns for each census tract i for North St. Louis was obtained by multiplying the overall burn rate for North St. Louis among children 0–14 years of age by the number of children aged 0–14 in census tract i obtained from the 1990 census. This method calculates how many burns were expected based on the population in a census tract if the burn rate for North St. Louis is applied to that census tract. Bayesian analysis relies on prior information. Selecting suitable priors is a key issue in spatial statistics since this prior information may strongly influence the posterior estimates of the relative risks.21 The results are based on the following prior distributions: G (0.12, 0.016) for the spatially structured component and G (5, 0.12) for the spatially unstructured component. The parameters of these prior distributions were based on the variance of the log (observed/expected burns) and appear reasonable in the absence of prior information about the relative importance of the spatially structured and unstructured components of the model.22 We used several different priors (both uninformative and informative priors) to determine the stability of the findings for the structured and unstructured errors, all of which followed the Gamma distribution, namely G (0.005, 0.005), G (0.5, 0.0005), G (0.001, 0.001), G (1,1), and G (5,5). Informative priors indicate that prior information about the variability of the burn rates was available. The model fit was determined based on the deviance, with lower deviance indicating better fit. Results of the Bayesian analysis are based on the posterior distribution, estimated using the Gibbs sampler under Markov Chain Monte Carlo, and implemented with the winBUGS software.23,24 Gibbs

sampling is an adaptation of the general Metropolis algorithm to simulate a Markov chain in Monte Carlo simulations. Gibbs sampling consists of visiting each parameter in the model in turn and simulating a new value for this parameter given the current values for the remaining parameters. Two separate chains starting from different initial values were run for each model. Convergence was checked by visual examination of ‘‘time series’’ style plots of the samples for each chain, and by calculating the Gelman and Rubin diagnostic.25 Based on this, the first 4,000 samples of each simulations were discarded as ‘‘burn-in.’’ Each of the two chains was run for an additional 20,000 iterations and posterior estimates were based on pooling the 40,000 samples for each model. Calculation of the model required construction of an adjacency matrix. Census tracts were considered adjacent when they had portions of their boundaries in common. The adjacency weights were zero unless census tracts were geographically adjacent. For adjacent census tracts, weights were one. Thus, the adjacency matrix consists of rows and columns of the census tracts with their value being one if they were adjacent and zero if they were not. ArcView/GIS (version 3.2) showed that the average number of neighbors per census tract in North St. Louis was 5.4. This software was also used to thematically map the RRFB estimates. Data were grouped according to the Fisher-Jenks algorithm.26 With this method, the differences between data values in the same class are minimized and the differences between classes are maximized, and at the same time groupings and patterns inherent in the data set are better represented.

RESULTS Burns in the City of St. Louis. The city of St. Louis is roughly divided on a north–south gradient by Highway I-64. SLCH lies to the north of I-64 and CGH lies just to the south (Figure 1). There were 311 children from the city of St. Louis evaluated for burn injuries in the study hospitals in 1995. Mapping these injuries showed that North St. Louis had an incidence of burn injuries higher than the rest of St. Louis (Figure 3). The distributions of gender and age of Missouri, St. Louis, and North St. Louis patients treated for burn injuries were similar. The percentage of African American patients from North St. Louis (98%) was higher than for Missouri (72.2%) or St. Louis (85.2%). Age distributions, median education levels, and median housing ages were similar for St. Louis and North St. Louis. However, socioeconomic indicators were not comparable (Tables 1 and 2). Two types of maps are used in this study—one depicting injury rates and one depicting injury risk. Based on burn rates in children aged 0–14 years,

748

Williams et al.

TABLE 2. Socioeconomic and Demographic Characteristics of St. Louis and North St. Louis from the 1990 United States Census Bureau* St. Louis Total population Total population under 14 years of age

North St. Louis

d


Burns in North St. Louis. Next, we focused on North St. Louis for the following two reasons: 1) capture of the data for burns from the northern part of St. Louis City was higher, and 2) we identified an area where the burn rate was elevated. Burns identified in the Missouri State database, which has had mandated E-code reporting since 1993, was compared with burns identified from hospitals participating in our study (Figure 2). The number of burns reported by the study hospitals as a percent of burns reported to the state ranged from 44.4% to 114.3% for St. Louis and 42.9% to 97.4% for North St. Louis. Information reported to the state is by zip code. Hierarchical Bayesian analysis of North St. Louis showed evidence of spatial clustering of the burn rate. There were 196 burns in 58 census tracts in North St. Louis. Observed rates varied from a minimum of 0 to a maximum of 11.5 per 1,000, with a mean rate of 3.4 burn injuries per 1,000 children 0–14 years (SD 6 2.6) (Table 3). To identify areas where the risk of burns was higher than expected, we constructed a map of North St. Louis using census tracts as the unit of analysis. We chose census tracts as the unit of analysis because smaller geographic divisions are more likely to have some units without burns. Relative risk by census tract varied within North St. Louis, from 0.87 to 1.182 (Figure 5). The 95% credible intervals (95% CIs) for all relative risks spanned the value of one.

396,685

193,224

80,399 (20.2%)

42,366 (21.9%)

60.9%

93.9%

47.4% 51.0%

83.4% 15.5%

1.6%

1.1%

8.0% 7.2% 5.1% 5.0%

8.2% 7.7% 6.0% 5.9%

40.1%

49.2%

24.6% 39.7%

33.3% 49.7%

50.9% 10.9%

68.2% 15%

Median education of persons aged $25

11.95 yr

11.75 yr

DISCUSSION

Median household income

$19,458

$15,403

Median age of housing (year built)

1935

1940

Using a GIS to map filtered block-group burn rates, we identified an area in North St. Louis with elevated rates of pediatric burn injuries. Demographic information revealed that the percentage of African American patients with burns from North St. Louis (98%), was higher than from Missouri (72.2%) or the city of St. Louis (85.2%), reflecting the fact that the percentage of African American population aged 0–14 years is higher in North St. Louis (93.9%) than in the City of St. Louis (60.9%). St. Louis and North St. Louis were comparable in terms of percentage of total population less than 14 years of age, age distribution, median education of adults, and median age of housing. Socioeconomic indicators such as percent of children in poverty, percent unemployment, and median household income differed between St. Louis and North St. Louis. Although it is well known that many of these variables are interrelated, their relationship to burn risk is an important area for future study. Overall, we were able to account for approximately 77% of the St. Louis City and approximately 86% of the North St. Louis burn cases reported to the state database. Because zip codes span the borders of the city of St. Louis and North St. Louis, the percent capture rate is approximate. State injury databases have the potential to be used for geographic analysis and could increase what we currently know about

African American population under 14 years of age Race/ethnicity African American White Other (Native American, Eskimo, Aleut, Asian, other race) Age distribution 0–4 years 5–9 years 10–13 years 14–17 years Socioeconomic indicators Households earning less than $14,999/year Persons below poverty level Children in poverty Single-parent households with children Unemployed

*From: www.censusbureau.com and www.mcdc.missouri.edu.

generated from our study data, the incidence of injuries was higher in North St. Louis than in the remainder of St. Louis. Block-groups were used as the unit of analysis in order to achieve maximum resolution and identification of the smallest geographic area where injury rates were elevated. Figure 3 shows a high incidence of burn rates in North St. Louis as a contour with the number six. This represents the filtered rate contour of 6 burn injuries per 1,000 children aged 0–14 years at the line border. Within this area, block-group burn rates ranged from 0 to 58.8 per 1,000 children at risk. Although there are other regions depicted with similarly high blockgroup rates, this one was the largest area with this elevated filtered rate, and it is the only one found where there was a relatively high population density (Figure 4). There are three other areas of the map with burn rate contours of 6 per 1,000 children aged 0–14 years, but these are in areas where there are few children or where there is an edge effect related to the mapping software.

ACAD EMERG MED

d

July 2003, Vol. 10, No. 7

d

749

www.aemj.org

Figure 4. 1990 population density by census tract.

local and regional injury patterns. However, in Missouri, the information is available only at a zip code level of analysis. Zip codes usually cover large areas, are socioeconomically heterogeneous, and are not consistent over time because their assignment is at the discretion of postal officials. Therefore, it is difficult to construct detailed maps based on zip code data that are reliable over time. For mapping at the block-group level, one needs to know the addresses of individual patients. Currently, this requires abstracting data from medical records, which is a slow, laborintensive process. Unintentional injuries are a significant health issue among children.1,2,15 Traditional surveillance methods have focused on the epidemiology and ecological correlates of childhood injury. Such studies have helped to characterize childhood injury patterns and

factors that may be associated. Although simple mapping strategies have been used to study injury, to our knowledge, there have been no published reports using a combination of GIS and spatial statistics to examine burn injuries. This study shows that geographic mapping techniques and spatial statistics are a useful addition to pediatric injury surveillance methods. Just as rates over time are less variable than the rate at a given point in time, rates averaged over neighboring areas (space) are more stable than the rates for individual units of geography. Using a GIS to map filtered blockgroup burn rates, we were able to identify an area of the city with elevated rates of pediatric burn injuries. A hierarchical Bayesian approach to mapping was then employed to determine the extent of local geographic variability of the burn rates. This

750

Williams et al.

d


TABLE 3. Socioeconomic and Demographic Data for St. Louis from 1995 Pediatric Burn Data N* Children aged under 15 years % Nonwhite population Median grade level for persons aged over 25 years Median household income for 1989 % Persons unemployed Median age of housing % Children in poverty % Children in single-parent household Burn rate per 1,000 at risk

Mean

SD

Minimum

Maximum

113 112

754 50.5%

414 40%

0 0.13%

2,120 100%

112 112 111 111 111 111 111

11.9 $18,404.84 12% 55.6 yr 35.4% 48.6% 3.4

0.99 $6969.47 7.7% 7.9 yr 22.3% 26% 2.6

9.4 0 0 25 yr 0 0 0

15.5 $33,549.00 35.8% 60 yr 85% 97.7% 11.5

*N ¼ number of census tracts: total 113 census tracts from 1990 census; 111 census tracts had children living in them; 112 census tracts had population in them.

approach takes into account that geographically close areas may have more similar burn rates than geographic areas that are further apart. The use of priors is an important aspect as well in Bayesian analysis, since it ensures that data are not interpreted in isolation from previous knowledge. Although we found that there were some census tracts within North St. Louis with RRFB greater than one, suggesting that the North St. Louis areas did contain some

Figure 5. 1995 St. Louis City pediatric burns—relative risk from hierarchical Bayesian analysis. Note: Relative risks are displayed and have been grouped according to the Fisher Jenks algorithm.26 The 95% confidence intervals for all relative risks include 1.

areas of increased risk, the 95% CI for these census tracts included unity.

LIMITATIONS Although injury rates generated from contours based on filtered rates tend to be more stable than those calculated from data analyzed without these techniques, the current study is based on a limited number of observations. A single year of information is not sufficient to calculate a stable rate of injury. Additional data would enhance the study by narrowing the credible intervals and, perhaps, decreasing the size of the high-risk areas. A stable rate is important for several reasons. First, identification of a geographic area allows for correlation with existing social and demographic data. This process may help to identify a population or factor that may contribute to the elevated rate.6 Prevention strategies can then be targeted to the specific geographic area, population, or factor. If one were to use an unreliable rate to plan prevention strategies, these efforts might be misdirected. Additionally, if the rates of burn injuries decreased in the given areas after prevention programs were initiated, it would be unclear whether the decrease in injury rates was due to the programs or the natural fluctuation in unstable rates over time. We believe that increasing the size of the database would provide a unique picture of the geographic location of burn injuries in the city of St. Louis. This would be important for injury prevention planning and for studying the effects of injury prevention efforts. A second issue, common to epidemiologic injury research, is that the location where the injury occurred may differ from the patient’s home address. Information regarding where an injury actually occurred is not routinely obtained. Therefore, maps depicting injury patterns in a geographic area usually are based on the location of residence. Inappropriate conclusions regarding the individual’s relationship to injury based on ecological data are a risk in ecological studies.27 If the conclusions are based on information that is not representative of the location where the

ACAD EMERG MED

d

July 2003, Vol. 10, No. 7

d

751

www.aemj.org

injury occurred, this potential may be magnified. However, currently the patient’s home address is the most consistently obtained information available for geographic-based studies. A prospective study using a GIS to map actual injury location versus the patient’s home address might provide valuable insight into this issue. The proximity of the two major pediatric hospitals to each other and to the geographic area of highincidence burn injury raises the question of referral bias. We do not have data for the community hospitals to the south and west of the area upon which we focused our analysis. However, we were able to identify approximately 77% of the burn cases reported to the State of Missouri for St. Louis and 86% for North St. Louis. This suggests that the two study hospitals collectively account for the majority of the pediatric burn patients cared for in the City of St. Louis, making it unlikely that the high incidence area would be greatly altered by the additional cases. However, only injuries treated in an emergency unit or admitted to the hospital were studied. As injuries treated in physicians’ offices or clinics are not reported to the state, we do not know the number or geographic location of these injuries. Hence, we do not know whether the addition of these data would change the area we identified as high-risk. Finally, we used small area injury data to generate maps. Injury risk in an area with a small population denominator could be influenced by a few high-risk individuals or families or by population changes in the area. However, the use of filtered contours and spatial statistics helps to reduce fluctuation due to small numbers by taking into account the injury rates in neighboring areas. Despite these limitations, there are several potential uses for GIS technology within the field of injury research. Mapping injuries to a small geographic area provides a vivid picture of at-risk populations. Once discrete geographic locations are documented, small pockets within them can be identified and targeted for injury prevention initiatives. Further research using this information may provide clues as to why certain areas are at higher risk than others. This might occur through analysis of factors unique to a geographic area, such as a paucity of stoplights in an area with a high incidence of pedestrian injury. Analysis of existing demographic data, such as that available from the U.S. Census, might also provide information. A high incidence of house fires in an area with old houses might suggest faulty wiring or lack of smoke alarms as potential culprits. Detailed maps of specific injury patterns could be used to focus prevention efforts toward high-risk geographic areas and to monitor injury patterns over time. This type of information could be important to those involved in community planning and education. If an area or population at high risk for injury is identified, injury prevention and

education programs can be more specifically tailored and injury rates can be tracked over time.

CONCLUSIONS The use of GIS and spatial statistics is an important addition to the study of childhood injuries. The value of this combination of methods is that small data sets can be used to generate relatively stable injury rates, because the data are filtered, and because the method takes into account what is happening in neighboring areas. Using these techniques, we were able to depict the geographic pattern of burn injuries in a major metropolitan area. The approach we used provided a detailed level of surveillance, allowing for identification of small geographic areas with elevated rates and risks of one particular type of injury. The ability to identify census tracts or block-groups at risk for injuries is an important surveillance technique that could serve as a springboard to injury prevention efforts. A map is a clear way to communicate information. Maps that illustrate patterns of injury in a city may be of use to those involved in health planning and allocation of resources or to those who appeal to them for funding for prevention initiatives. Helping communities recognize their increased risk of particular injuries may serve to increase interest in local prevention efforts. We hope that the use of GIS and spatial statistics will aid in the description of injuries and that this information subsequently may be used to develop targeted intervention strategies. References 1. Office of Statistics and Programming, National Center for Injury Prevention and Control, CDC. Webstite: www.cdc.gov/ ncipc/wisqars. Accessed March 2002. 2. Rivara FP, Grossman DC. Prevention of traumatic deaths to children in the United States: how far have we come and where do we need to go? Pediatrics. 1996; 97:791–7. 3. Cooper A, Barlow B, Davidson L, Relethford J, O’Meara J, Mottley L. Epidemiology of pediatric trauma: importance of population-based statistics. J Pediatr Surg. 1992; 27:149–54. 4. Marganitt B, MacKenzie EJ, Deshpande JK, Ramzy AI, Haller JA. Hospitalizations for traumatic injuries among children in Maryland: trends for incidence and severity: 1979 through 1988. Pediatrics 1992; 89:608–13. 5. Braddock M, Lapidus G, Gregorio D, Kapp M, Banco L. Population, income, and ecological correlates of child pedestrian injury. Pediatrics. 1991; 88:1242–7. 6. Rivara FP, Barber M. Demographic analysis of childhood pedestrian injuries. Pediatrics. 1985; 76:375–81. 7. Istre GR, McCoy MA, Osborn L, Barnard JJ, Bolton A. Deaths and injuries from house fires. N Engl J Med. 2001; 344:1911–6. 8. Mallonee S, Istre GR, Rosenberg M, et al. Surveillance and prevention of residential-fire injuries. N Engl J Med. 1996; 335:27–31. 9. Hazinski MF, Eddy VA, Morris JA. Children’s Traffic Safety Program: influence of early elementary school safety education on family seat belt use. J Trauma. 1995; 39:1063–8.

752 10. Clamp M, Kendrick D. A randomized controlled trial of general practitioner safety advice for families with children under 5 years. BMJ. 1998; 316:1576–9. 11. Rivara FP, Thompson DC, Thompson RS, The Seattle Children’s Bicycle Helmet Campaign: changes in helmet use and head injury admissions. Pediatrics. 1994; 93:567–9. 12. Luria JW, Smith GA, Chapman JI. An evaluation of a safety education program for kindergarten and elementary school children. Arch Pediatr Adolesc Med. 2000; 154:227–31. 13. Duperrex O, Bunn F, Roberts I. Safety education of pedestrians for injury prevention: a systematic review of randomized controlled trials. BMJ. 2002; 324:1129–33. 14. Nigel P, Vinten-Johnasen P, Brody H, Rip M. A rivalry of foulness: official and unofficial investigations of the London cholera epidemic of 1854. Am J Public Health. 1998; 88:1545–53. 15. Quayle KS, Wick NA, Gnauck KA, Schootman M, Jaffe DM. Description of Missouri children who suffer burn injuries. Inj Prev. 2000; 6:255–8. 16. Rushton G, Krishnamurthy R, Krishnamurti D, Lolonis P, Song H. The spatial relationship between infant mortality and birth defect rates in a U.S. city. Stat Med. 1996; 15:1907–19. 17. Rushton G, Lolonis P. Exploratory analysis of birth defect rates in an urban population. Stat Med. 1996; 15:717–26. 18. Rushton G, Armstrong MP. Improving public health through geographical information systems: an instructional guide to major concepts and their implementation [CD-ROM]. IA: University of Iowa, Iowa City, 1997.

Williams et al.

d


19. Lewis RJ, Wears RL. An introduction to the Bayesian analysis of clinical trials. Ann Emerg Med. 1993; 22: 1328–36. 20. Besag J, York J, Mollie A. Bayesian image restoration with two applications in spatial statistics. Ann Inst Stat Math. 1991; 43:1–59. 21. Bernardinelli L, Clayton D, Montomoli C. Bayesian estimates of disease maps: how important are priors? Stat Med. 1995; 14:2411–31. 22. Mollie A. Bayesian mapping of disease. In: Gilks WR, Richardson S, Spiegelhalter DJ (eds). Markov Chain Monte Carlo in Practice. Boca Raton, FL: Chapman & Hall, 1996, pp 359–79. 23. Wakefield J, Best N, Waller L. Bayesian approaches to disease mapping. In: Elliot P, Wakefield J, Best N, Briggs D (eds). Spatial Epidemiology. Oxford: Oxford University Press, 2000, pp 104-27. 24. Spiegelhalter D, Thomas A, Best N, Gilks W. BUGS 0.5* Bayesian inference using Gibbs Sampling Manual (version ii). 1996; Cambridge, England. 25. Gelman A, Rubin DB. Inference from iterative simulation using multiple sequences. Stat Sci. 1992; 7:457–511. 26. Slocum TA. Thematic Cartography and Visualization. Upper Saddle River, NJ: Prentice Hall, 1999, p 239. 27. Kelsey JL, Whittemore AS, Evans AS, Thompson WD. Methods in Observational Epidemiology, 2nd Edition. New York, NY: Oxford University Press, 1996, pp 262–3.