A Procedure for Detecting Childhood Cancer Clusters Near Hazardous Waste Sites in Florida
Posted on: Friday, 2 May 2008, 06:01 CDT
By Kearney, Greg
Abstract Despite over 20 years of research on childhood cancer clusters and hazardous waste sites, little evidence has been produced to indicate a causal relationship. Nevertheless, the perception of a childhood cancer cluster being located near a hazardous waste site can raise fear and uncertainty, and it demands attention from health officials. To investigate this public health concern, the author used the spatial-scan statistical software SaTScan to detect childhood cancer clusters and their proximity to National Priority List (NPL), or Superfund, sites in Florida. In the ecological study reported here, "most likely" clusters were defined as those with a p-value of <.05. Distance served as a proxy for exposure; a geographical information system (GIS) was used to determine the number of clusters within a predetermined distance of an NPL site. Spatial clusters were found to occur randomly throughout the state, with most clusters being identified in the more populated counties, and clusters less likely to occur near an NPL site. This article attempts to explain the utility of an emerging public health surveillance tool for detecting cancer clusters near hazardous waste sites. Despite several epidemiological limitations of the study, as well as the fact that there are other environmental exposure hazards such as Toxic Release Inventory facilities and landfills, the SaTScan program proved useful as a surveillance tool for generating more in-depth studies.
Introduction
For obvious reasons, a childhood cancer cluster located near a hazardous waste site is a serious public health concern. Despite extensive studies conducted over the past 20 years, little supporting evidence has been found to suggest an association, and the etiology of childhood cancers remains relatively unknown (Caldwell, 1990). Nevertheless, the perception of living near a hazardous waste site can raise fear and concerns. According to a Princeton survey of 801 registered voters in Florida, 89 percent felt that environmental factors such as pollution and toxic wastes played a significant role in causing diseases. Approximately 34 percent of parent voters with young children felt that childhood cancers such as leukemia were linked to environmental factors such as hazardous waste sites (Princeton Survey Research Associates, 2000). In addition, cluster inquiries usually involve local or state public health offices and require them to respond. Researchers at the Johns Hopkins Bloomberg School of Public Health found that 95 percent of state health departments surveyed (N = 37) reported having a cancer cluster investigation or inquiry. Of the states included in the survey, 81 percent reported brain cancer as the most common type of cancer cluster inquiry followed by leukemias, lymphomas, or both (72 percent) (Juzych et al., 2007).
To explore the issue of cancer clusters near hazardous waste sites, the author used a geographic information system (GIS) and spatial-scan statistical software, SaTScan (Version 6.0) to spatially and temporally detect the geographical location of childhood cancer clusters near U.S. Environmental Protection Agency (U.S. EPA) National Priority List (NPL) sites in Florida (U.S. EPA, n.d.). The NPL sites were chosen because they are considered the worst-of-the-worst hazardous sites in the nation.
Background
SaTScan is a software tool developed by the National Cancer Institute for cluster detection and can be downloaded at http://www satscan.org/. SaTScan can work by any of the following methods: a) evaluate reported spatial or space-time disease clusters and determine if they are statistically significant, b) test whether a disease is randomly distributed over space, time, or space-time, c) perform geographical surveillance of disease, d) test geographical areas of significantly high or low rates, and e) perform repeated time-periodic disease surveillance for the early detection of disease outbreaks (NCI, 2004). The program uses a circular window filter that scans over a map, smoothing out the data while analyzing for clusters. The circular windows are generated around geographical centroids, or central longitude/latitude coordinates of polygons projected on a map. Using this software, the author detected childhood cancer clusters and estimated their proximity to an NPL site using a geographic information system (GIS).
Methods
Cancer case records were identified from confidential data retrieved with Institutional Review Board (IRB) approval from the Florida Cancer Data System (FCDS) (FCDS, n.d.). The study period was January 1, 1990, through December 31, 1999. The case definition included any resident of Florida who was between O and 19 years of age at the time of diagnosis and who was diagnosed with a new primary cancer or reportable condition during the study period. The childhood cancer cases included those with a primary diagnosis of selected leukemias, lymphomas (not including Hodgkin's lymphoma), brain cancer, or central nervous system (CNS) cancer. As indicated in Table 1, the cancer types were determined by the International Classification of Diseases code (ICD-0-02) and the International Agency for Research on Cancer, International Classification of Childhood Cancers histology codes (IARC, 2003).
A comprehensive list of NPL sites was located on the U.S. EPA Web site (http://www. epa.gov/superfund/sites/query/basic.htm). The list included the names, addresses, longitude and latitude coordinates, and site status of 70 NPL sites in Florida. Originally, 58 NPL sites were selected for the study on the basis of their public health threat status as determined by the Agency for Toxic Substances and Disease Registry (ATSDR). Ten sites were removed from consideration, however, because they were neither "currently on the final NPL" nor "proposed for the NPL" by US. EPA from 1985 to 2003. Each of the geographic information system (GIS) shape files that were used to provide the mapping for the project were supplied by the Florida Data Geographical Library (FDGL), the Florida Department of Environmental Protection (FDEP), or the Florida Department of Health (FDOH). The annual population size of each census tract was determined by linear interpolation and extrapolation of the 1990 and 2000 population data from the U.S. Census Bureau (2002a, 2002b).
Spatial Distribution of Childhood Cancer Clustering-SaTScan
SaTScan analyzed the spatial distribution of each of the three cancer site groupings (leukemias, lymphomas, and brain/CNS cancers) and the combined cancer groupings at the geographic census tract level for clustering. A relative risk was computed on the basis of the observed and expected numbers of cases for each census tract. Under the Poisson model, the expected number of cases under the null hypothesis was calculated for each census tract by means of the indirect standardization method, without covariate adjustment (i.e., for age, sex, race), as follows:
E[c] = p * C/P
Where
c is the observed number of cases, p is the population in the location of interest, and
C and P are the total numbers of cases and population, respectively (Kulldorff, 2004).
A log likelihood ratio was calculated for each region to test whether the risk in the area under the scanning window was elevated by comparison with the risk in the area outside the window, which would indicate the existence of a cluster of cases.
Under the Poisson assumption, the log likelihood function for a specific window is proportional to the following:
Where
C is the total number of cases, c is the observed number of cases within the window, and
E[C] is the covariate adjusted expected number of cases within the window under the null hypothesis.
The analysis is conditioned on the total number of cases observed, C - E[c], being the expected number of cases outside the scanning window. I( ) is the indicator function. With SaTScan set to scan for clusters with high rates, I( ) is set to 1 when the window has more cases than expected under the nullhypothesis; it is set to O otherwise. When SaTScan is detecting for clusters with either high or low rates, as in this application, the indicator function is I( ) = 1 for all windows (Kulldorff, 2004).
From the FCDS raw-data file, all duplicated cancer cases were removed, and the remaining cases formatted into a Microsoft Excel spreadsheet. Next, the geographical location of each case was identified and sorted by census tract and cateorgorized into the following groups:
1. selected leukemia cancers, ages 0 to 19;
2. selected lymphoma cancers, ages 0 to 19;
3. selected brain and CNS cancers, ages 0 to 19; and
4. combined cancer groupings (selected leukemias, selected lymphomas, and brain and CNS cancers), ages 0 to 19.
As indicated in Table 2, the SaTScan spacetime analysis requires three data files to properly run the analysis: 1) a population file, 2) a case file, and 3) a geographic-coordinates file. The population file contained the actual population count for each area obtained from the 1990 and 2000 U.S. censuses. The case file contained the actual number of childhood cancer cases for each specified census tract, and the date of diagnosis for each case. The geographic coordinates file represents an area that can be aggregated and represented by a single geographical point location, for example the latitude and longitude coordinates of a city, county, or census tract. Here, the geographical centroids (latitude and longitude coordinates) of each of the 3,154 census tracts in Florida were chosen. If a cluster was detected by SaTScan, it is recognized by the geographical boundary surrounding the geographical centroid of each census tract. To further explain: The SaTScan space-time scan statistic is characterized by a cylindrical scanning window (or circle), which moves across the study region, circling each of the 3,154 census tracts centroids in which no more than 50 percent of the population is at risk. If the census tract centroid falls within the scanning window and is considered covered by the window, then SaTScan computes the numbers of cases inside and outside each circle. If SaTScan detects a cancer cluster within the scan analysis window with a probability value (p-value) of
The cylinder that has the highest calculated maximum log likelihood ratio (LLR) and has more than the expected numbers of cases is considered the "most likely" cluster. Other detected clusters are considered "secondary" clusters and are evaluated on the basis of the significance of their p-value.
Using SaTScan, the author analyzed each of the three cancer types and the total cancers on a Dell computer with a Pentium M processor and the following specifications: 1,600 MHz, 1.60 GHz, and 1.00 GB of RAM.
Cluster Detection and Proximity to NPL Sites
Using the buffer tool in ArcMap (Version 8.1) (LSRI, 2004), a 0- to-1-mile, l-to-5-mile, and 5-to-10-mile radii were selected around each of the 48 NPL sites. Next, the author mapped the location of each of the "most likely" statistically significant (p
Results
During the 10-year study period, a total ol 1,125 cases of leukemias, 126 cases of nonHodgkin's lymphomas, and 633 cases of brain and CNS cancers were identified. When the cancer groupings were combined and spatially analyzed, 1,884 cancers were selected for a total crude incidence rate of 43 cases per 100,000. Before exploring the space-time analysis, a "purely spatial" analysis for each of the cancer groupings was performed, followed by the combined cancer groupings, to examine the location of the "most likely" clusters. As indicated in Table 3, most of the spatial clustering was identified in Broward and Miami-Dade counties.
As indicated in Table 4, when a temporal component was added to the spatial analysis, making it a space-time analysis, the locations of the clusters that were detected were similar to those found by the purely spatial analysis. The clusters identified as "most likely" (p = .001) were leukemias and total-cancer clusters in Hillsborough County, with seven cases observed and 0.09 cases expected, and 8 cases observed and 0.16 expected, respectively. The "most likely" (p = .023) lymphoma clustering (5 cases observed, 0.11 expected) and brain/CNS clustering (13 observed and 0.09 expected) were detected in Miami-Dade County.
Clustering and Proximity to NPL Sites
As shown in Figure 1, the spatial scan detected the "most likely" (p
Miami-Dade County, which shares a common north-south border with Broward County, had 62.5 percent more NPL sites within 5 to 10 miles of a detected cluster than within 0 to 1 mile, where 12.5 percent of sites were located. In central Florida, Polk County had one NPL site within 0 to 1 mile of a detected cluster. "Secondary" clusters, where the LLR ranked lower and p < .05, were detected in Hillsborough and Polk Counties, but no NPL site was located within a radius of 0 to 1 mile.
Discussion
Among the major issues surrounding a study of this nature is the need to recognize that the etiology of childhood cancers is still relatively unknown. A confirmed cancer cluster could be the result of chance, miscalculation of the expected number of cancer cases (e.g., failure to consider a risk factor within a population at risk), differences in case definition for observed cases and expected cases, known causes of cancer (e.g., smoking), or some other unknown cause (CDC, 2002). An estimation of whether an NPL site had an adverse health effect on a population can be reached only through more rigorous studies.
Limitations of the study include the use of the geographical centroids of each of the Florida census tracts as cluster boundaries. The health data used at the census tract level were aggregated to the geographic centroid of each census tract and may not reflect the real cluster location of the cluster; therefore, only an estimate of the position and an estimate of the radius of the cluster are provided. Most of the clusters detected within a 1- mile radius of an NPL site were in proximity to six NPL sites in Miami-Dade and Broward Counties. The areas in which these clusters were identified are characterized by mixed urban, commercial, residential, and industrial uses. Some of these areas could be considered densely populated, and without more specific geographic data, it is difficult to determine with precision where the clusters actually exist.
Also, public health assessment reports conducted by the Florida Department of Health and ATSDR on these sites found evidence of soil contamination, groundwater contamination, or both from volatile organic compounds, heavy metals, and polychlorinated biphenyls (PCBs). However, no evidence of human exposure was reported,
As a surveillance tool, SaTScan has additional limitations. Because childhood cancers are rare, a few observed cases may be statistically significant on the basis of the calculated number of expected cases. Having a small number of observed cases and a low number of expected cases may not necessarily mean that a census tract contains a significant cluster; as previously mentioned, this outcome could be the result of chance alone. Another important epidemiological limitation is the geographic uncertainty and difficulty of estimating time and place of exposure. The diagnostic address of the child may not necessarily reflect where a child was living at the time of exposure.
The rationale behind SaTScan is detecting disease clusters in the population from data provided by the user. The program offers several advantages for the identification of childhood cancer clusters: 1) it does not assume a priori knowledge, which can lead to pre-selection bias; 2) it produces analytical results with a file extension that makes the data easily imported into a GIS; 3) it adjusts for non-homogeneous population density; 4) it allows for covariate adjustments that control for confounders;, and 5) it provides a single p-value for testing the null hypothesis. In addition, covariate adjustments such as adjustments for age and race could be included in the SaTScan analysis and could serve as a springboard for more in-depth environmental health studies surrounding NPL sites.
REFERENCES
Caldwell, G.G. (1990). Twenty-two years of cancer cluster investigations at the Centers for Disease Control. American journal of Epidemiology, 132(Suppl. 1), 43-47.
CDC, Department of Health and Human Services. (2002, October). Cancer clusters, (Fact sheet). Retrieved August 24, 2003, from http:/ /www cdc.gov/nceh/clusters/cncr_clstr_fctsht.pdf
ESRI. (2004). ArcMap (Version 8.1) [Computer software]. Redlands, CA: Author.
Florida Cancer Data System (FCDS). (n.d.). Unpublished raw cancer data [Data file].
International Agency for Research on Cancer, (n.d.). International classification of childhood cancer (ICCC) from IARC technical report No. 29. Retrieved November 23, 2003, from http:// seer.cancer.gov/iccc.
Kulldorff, M. (2004). SaTScan User guide for version 5.0. Retrieved October 23, 2004, from http://www.satscan.org/ techdoc.html.
NCI (2004). SaTScan Version 2.0 [Computer software and manual]. Retrieved March 23, 2004, from http://surveillance.cancer.gov/ statistics/types/ incidence.html.
Princeton Survey Research Associates. (2000, July). National survey of public perceptions of environmental health risks, Florida component (PSRA for Health-Track). Princeton, NJ: Author.
Juzych, N., Resnick, B., Streeter, R., Herbstman, J., Zablotsky J., Fox, M., & Burke, T.A. (2007). Adequacy of state capacity to address noncommunicable disease clusters in the era of environmental public health tracking. American Journal of Public Health, 97, (Suppl. 1), S163-S169.
U.S. Census Bureau. (2002a, modified September 2007). Census '90. Retrieved February 17, 2004, from http://www.census.gov/main/ www/ cen1990.html.
U.S. Census Bureau. (2002b, modified September 2007). United States Census 2000. Retrieved February 4, 2008, from http://www census.gov/main/www.cen2000.html.
U.S. Environmental Protection Agency, (n.d.) National priorities list: Basic query form. Retrieved September 23, 2003, from http:// www epa.gov/superfund/sites/query/basic.htm
Greg Keamey, Dr.P.H., M.P.H., R.S.
Acknowledgements: The author would like to express special thanks to Dr. Martin Kulldorff, Harvard University, and to the Florida Department of Health, Division of Environmental Health. Disclaimer: The work reported here was an independent research study and does not reflect the opinions or views of the Florida Department of Health.
Corresponding author: Greg Kearney Environmental Epidemiologist, Florida Department of Health, Division of Environmental Health, Bald Cypress way, Bin A08, Tallahassee, FE 32399. E-mail: greg_kearney@doh. state.fl.us.
Copyright National Environmental Health Association May 2008
(c) 2008 Journal of Environmental Health. Provided by ProQuest Information and Learning. All rights Reserved.
Source: Journal of Environmental Health
User Comments (0)
RSS Feeds