Anda di halaman 1dari 8

Pediatr Blood Cancer 2010;54:511518

Epidemiologic Mapping of Florida Childhood Cancer Clusters


Raid Amin,

PhD,

1,2

* Alexander Bohnert,1 Laurens Holmes, PhD, DrPH,2,3 Ayyappan Rajasekaran,


and Chatchawin Assanasen, MD2,4**

Background. Childhood cancer remains the leading cause of


disease-related mortality for children. Whereas, improvement in care
has dramatically increased survival, the risk factors remain to be fully
understood. The increasing incidence of childhood cancer in Florida
may be associated with possible cancer clusters. We aimed, in this
study, to identify and confirm possible childhood cancer clusters and
their subtypes in the state of Florida. Methods. We conducted purely
spatial and spacetime analyzes to assess any evidence of childhood
malignancy clusters in the state of Florida using SaTScanTM. Data
from the Florida Association of Pediatric Tumor Programs (FAPTP) for
the period 20002007 were used in this analysis. Results. In the
purely spatial analysis, the relative risks (RR) of overall childhood
cancer persisted after controlling for confounding factors in south
Florida (SF) (RR 1.36, P 0.001) and northeastern Florida (NEF)

Key words:

PhD,

(RR 1.30, P 0.01). Likewise, in the spacetime analysis, there


was a statistically significant increase in cancer rates in SF
(RR 1.52, P 0.001) between 2006 and 2007. The purely spatial
analysis of the cancer subtypes indicated a statistically significant
increase in the rate of leukemia and brain/CNS cancers in both SF
and NEF, P < 0.05. The spacetime analysis indicated a statistically
significant sizable increase in brain/CNS tumors (RR 2.25,
P 0.02) for 20062007. Conclusions. There is evidence of spatial
and spacetime childhood cancer clustering in SF and NEF. This
evidence is suggestive of the presence of possible predisposing
factors in these cluster regions. Therefore, further study is needed to
investigate these potential risk factors. Pediatr Blood Cancer 2010;
54:511518. ! 2010 Wiley-Liss, Inc.

cancer cluster; childhood neoplasm; cluster analysis; epidemiology; florida

INTRODUCTION
Cancer remains the leading cause of disease-related death
among children in the United States despite progress in clinical trials
and significant improvements in survival rates [1]. Over the past
20 years in the United States, increases in the incidence of childhood
cancer have also been observed from 11.5 cases per 100,000
children in 1975 to 14.8 per 100,000 children in 2004 [2]. In 2009,
approximately 10,730 children under the age of 15 will be diagnosed
with cancer and about 1,480 are projected to die from the disease [3].
Despite the burden of childhood cancer and the many years of
epidemiologic investigations, its causes remain largely unknown
but have been linked in small percentages to certain genetic
predispositions and exposures to chemotherapy agents and ionizing
radiation [47]. A number of studies continue to examine the
complexities of other possible risk factors for childhood cancers [8
11]. These include early-life exposures to infectious agents;
parental, fetal, or childhood exposures to environmental toxins;
parental occupational exposures to radiation or chemicals; parental
medical conditions during pregnancy or before conception;
maternal diet during pregnancy; early postnatal feeding patterns
and diet; and maternal reproductive history [1224]. Environmental factors may play an important etiologic role in childhood
malignancies and can be evidenced by excessive numbers of cases in
a defined geographic area relative to other areas, termed clusters.
A cancer cluster can be defined as the occurrence of a greater
than expected number of cases of a malignancy within a group of
people, a geographic area, or a period of time. There exist various
definitions of the terms cluster and clustering in the context of
spatial epidemiology and cancer research, respectively [25,26].
Identification of spacetime variations in incidence rate patterns
can provide important clues for further in-depth studies into the
etiology and control of cancer [27]. Spatial clustering is defined as a
general irregular spatial distribution of cases that is not confined to
one particular small area. Spacetime cancer clustering is observed
when an excess number of cases occur within a geographical
location over very limited periods of time and cannot be explained
in terms of general excesses in these locations or time frames.
Regional, national, and international registries have been utilized to

! 2010 Wiley-Liss, Inc.


DOI 10.1002/pbc.22403
Published online 6 January 2010 in Wiley InterScience
(www.interscience.wiley.com)

investigate possible spatial and spacetime clustering and any


associated risks of cancer predisposition [2124,2831].
In Florida, overall cancer statistics are similar to the rest of the
United States. From 1981 to 2000, 10,238 new cases of cancer were
diagnosed among Florida children and adolescents, representing
0.7% of all cancer cases diagnosed in the state [28]. The Florida
Association of Pediatric Tumor Programs (FAPTP) was founded in
1970 as a statewide network of childrens cancer programs under the
auspices of the Florida Regional Medical Program (FRMP). The
Florida legislature established a pediatric hematology/oncology
program within Childrens Medical Services (CMS) and FAPTP that
was given the responsibility and authority to monitor and evaluate
pediatric cancer care statewide. This reporting system provides the
state and the public with data on cancer incidence, clinical trial
participation, and survivorship. The Statewide Patient Information
Recording System (SPIRS) registers patients from the 16 pediatric
hematology/oncology centers statewide. In addition, the Florida
Cancer Data System (FCDS) captures the data from patients treated
outside the FAPTP system and can be linked with SPIRS data to
study the larger patient data base.

Department of Mathematics and Statistics, University of West Florida,


Pensacola, Florida; 2Nemours Center for Childhood Cancer Research,
Wilmington, Delaware; 3Department of Orthopedics, Alfred I duPont
Hospital for Children, Wilmington, Delaware; 4Nemours Childrens
Clinic, Pensacola, Florida
Conflict of interest: Nothing to report.
Grant sponsor: Nemours Childrens Clinic, Pensacola; Grant sponsor:
Nemours Foundation; Grant sponsor: Caitlin Robb Foundation.
*Correspondence to: Raid Amin, Department of Mathematics and
Statistics, University of West Florida, 11000 University Parkway,
Pensacola, FL 32514. E-mail: ramin@uwf.edu
**Correspondence to: Chatchawin Assanasen, Nemours Center for
Childhood Cancer Research, Nemours Childrens ClinicPensacola,
5153 North 9th Avenue, Pensacola, FL 32504.
E-mail: cassanas@nemours.org
Received 8 September 2009; Accepted 17 November 2009

512

Amin et al.

The recently founded Nemours Center for Childhood Cancer


Research (NCCCR) has three of its oncology clinics in Jacksonville,
Orlando, and Pensacola, Florida as well as one in Wilmington,
Delaware. One of the initial goals of this center was to evaluate
pediatric cancer epidemiology data in the states of Florida and
Delaware. In 2008, the Delaware childhood cancer rates were
evaluated by NCCCR in collaboration with Delaware Department of
Health and Social Services for possible childhood cancer clusters.
This assessment failed to confirm clusters probably due to small
number of cases as well as absence of clusters. The current study was
initiated about 2 years ago in collaboration with the University of
West Florida. We sought to identify and confirm overall childhood
cancer clusters as well as to determine whether or not clusters could
be confirmed by cancer subtypes. We utilized the data from FAPTP
and modeled our analysis using SaTScanTM to test the following
null hypotheses: (1) The pediatric cancer rate of all cancer types
is randomly distributed over space in Florida from 2000 to 2007,
(2) The pediatric cancer rate of all cancer types is randomly
distributed over time and space in Florida from 2000 to 2007,
(3) The rates for specific pediatric cancer types are randomly
distributed over space in Florida from 2000 to 2007, and (4) The
rates for specific pediatric cancer types are randomly distributed
over time and space in Florida from 2000 to 2007.

MATERIALS AND METHODS


We conducted purely spatial and spacetime analyzes to assess
the evidence of childhood cancer clusters in the state of Florida
using SaTScanTM.

Study Area and Population


We identified 67 counties and 972 zip code areas in Florida in
the year 2000. While the clustering evaluations could have been
based on Florida counties, we decided to obtain more detailed
information by using the zip code areas. The statistical analysis used
in this study requires that the geographic information for each zip
code area be represented by some form of a centroid. To obtain the
geographical centroid of each zip code area and to create maps with
information on the cancer clusters, the geographical information
system ArcGIS was utilized. We used consistent geographical data
for zip code areas, Zip Code Tabulation Area (ZCTA), from the year
2000 from the Florida Geographic Data Library (FGDL). For zip
code areas that were created after 2000 with an identified cancer
case, we manually assigned in the software package ArcGIS a zip
code based on its position in the 2000 geographical data set. A
marginal number of cases for which we could not determine their
position relative to the 2000 zip code area file were discarded. The
study population included the entire population of children 0
19 years of age in the state of Florida during the time period 2000
2007. These included children with and without the diagnosis of a
childhood cancer. During this time, there were 4,591 cases of
pediatric cancer diagnosed, of which 1,254 (27%) had leukemia,
839 (18%) had brain/central nervous system (CNS) cancer, and 252
(5.5%) had lymphoma.

Data Sources
The data for this study were available from FAPTP, an existing
de-identified dataset, that is, publicly available. FAPTP has been
Pediatr Blood Cancer DOI 10.1002/pbc

shown to be a valid and reliable source for pediatric cancer incidence


data in Florida [28,32,33]. The dataset included information on
cancer cases such as the diagnosis code for the study period 2000
2007 designated by the International Classification for Childhood
Cancer (ICCC) [34], incorporating the new codes introduced in
ICD-O-2 and the updated ICD-O-3. Demographic information was
also included, such as date of birth, age at cancer diagnosis, sex, and
zip code of residence. This study involved age-adjusted data. We
obtained Florida demographic population data such as age and race/
ethnicity from the 2000 census. For each ZCTA, we obtained the
total population at risk, stratified by age, sex, and race/ethnicity.

Data Analyzes
Clusters have been analyzed previously using several statistical
and epidemiologic approaches [35]. In this study, we used
SaTScanTM. The software package SaTScanTM [26] uses spatial
scan statistics to identify and test for the significance of cancer
clusters. The incidence counts in each zip code area are used either
in two dimensions for a purely spatial analysis or in a threedimensional setting for a spacetime analysis with the additional
dimension representing time. We assumed that the incidence of
cancer in each zip code area is distributed according to a Poisson
model [36,37]. This method tests the null hypothesis that the ageadjusted risk of cancer incidence is the same for all zip code areas.
With the covariates included in the model, we tested the null
hypothesis that within any age group, the risk of cancer incidence is
the same for the entire area covered in this study [37]. To include the
effect of urbanicity in our analysis, we used the population density
information for postal code level [37] that is available through the
Florida Geographic Data Library (FGDL). Possible associations to
socioeconomic status (SES) were investigated by using the
economics wealth index by Woods & Poole Economics Inc., which
we obtained from the HAAS Business Center at the University of
West Florida [38]. Since neither of these two covariates resulted in
any changes in the SaTScanTM computations, results on population
density or the socioeconomic status are not presented.
The spatial scan statistics in SaTScanTM identifies clusters by
imposing a window that moves over a map, including different sets
of neighboring zip code areas represented by their corresponding
centroids [29]. If the window includes the centroid of a specific zip
code area, then this zip code area is included in the window. As
suggested by Kulldorff et al. [29], the center of the window is
positioned only at the 972 zip code centroids. For each window, the
spatial scan statistic tests the null hypothesis of equal risk of
childhood cancer incidence for all zip code areas against the
alternative hypothesis that there exists an elevated risk of childhood
cancer incidence within the scan window when compared with areas
outside the window. The likelihood function for the Poisson model
can be shown to be proportional to
! n "n # N " n $N"n
I n > E
E
N"E
where n is the number of cancer incidences within the scan window,
N is the total number of incidences in Florida, and E is the expected
number of cancer incidences under the null hypothesis [29,37].
Since we are using a one-tailed test that rejects the null hypothesis if
there exists elevated cancer risk, an indicator function I is used such
that I 1 when the scan window has a larger number of cancer
incidences than expected if the null hypothesis were true, and zero

Florida Childhood Cancer Clusters


otherwise. It can be shown that for a given N and E, the likelihood
increases as the number of incidences, n, increase in the scan
window. How the spatial scan statistic within SaTScanTM actually
identifies cancer clusters is described elsewhere in detail [37]. By a
Monte Carlo simulation, we generated 999 random replications of
the data set to obtain the statistical stability for the identified cancer
clusters in the program SaTScanTM. The Monte Carlos test also
allows for the simultaneous controlling of multiple confounders
such as age, sex, race, income level, etc. The identified cancer
clusters are listed by SaTScanTM in order of significance such that
the P-value for each cluster is compared with a pre-set significance
level of 0.05.
There exist different types of the spatial scan statistics. Circular
or elliptical windows can be used to identify circular clusters and
elliptical shaped clusters, respectively. Both approaches were used,
and we arrived at virtually identical cluster results. In this study, we
present only the cancer clusters identified by circular windows.
While the spatial scan statistic requires specification of the
underlying distribution of the data used in SaTScanTM, making it
a parametric statistical method, a non-parametric smoothing
method was also used to check whether similar or identical cluster
results would be obtained. In particular, we used a weighted Head
Banging algorithm based on median smoothing which removes the
background noise of random variability so that the underlying
spatial pattern becomes more clear [3941]. Both parametric and
non-parametric methods were used for the purpose of results
validation. In this study, Head Bang was used to statistically double
check the results from SaTScanTM by removing local variations in
cancer incidence age-adjusted rates for the 972 zip code areas. This
particular smoother retains the important features, such as edges, but
smoothes out unreliable data points and spikes for low population
areas based on the chosen weights in the algorithm. To ensure
adequate statistical power, all cancer cases for the period 2000
2007 were used to perform a purely spatial analysis. For the space
time analysis, which is a temporal extension of the spatial analysis,
the algorithm searches within 20002007 for time periods in which
clusters appear.

RESULTS
The SaTScanTM purely spatial analysis of the FAPTP data
revealed two significant clusters in the state, one in southern Florida
and the other in northeastern Florida (NEF). The south Florida (SF)
cluster encompasses the southwest, south central and southeast
regions. The NEF cluster incorporates areas of the northeast and
north central regions. After adjusting for age, sex, and race as
covariates, a total of 4,181 cases were identified with a corresponding incidence rate of 14.4 average annual cases per 100,000. In SF,
there were 465 observed cases and 352 expected cases, with a
relative risk of 1.36, implying that compared with the state, there is a
statistically significant 36% increased risk of childhood cancer
(P 0.001). In the NEF cluster, there were 466 and 375 observed
and expected cases, respectively. This region appears to be smaller
in size, although it may represent a more densely populated area. A
similar increase in the rates of childhood cancer was identified with a
RR 1.30 (P 0.01). In addition, a third overall childhood cancer
cluster was identified in a small area of central Florida in which the
observed number of cases was 31 as compared to 11 expected cases.
The rates were statistically significantly higher in this area relative
to the state with a RR 2.82 (P 0.008), which implies that
Pediatr Blood Cancer DOI 10.1002/pbc

513

Fig. 1. Purely spatial analysis of FAPTP database for all cancer types
20002007. Clustering representation of SaTScanTM purely spatial
analysis is illustrated utilizing zip code data with age, sex, and race as
covariates. Clusters are represented in colors. The red area represents the
South Florida cluster. SaTScanTM computed results include: Coordinates/radius (26.3N, 81.3W)/101.6 km, Population 294,119,
Observed cases 465, Expected Cases 352. The orange area
represents the North Central Florida cluster. SaTScanTM computed
results include: Coordinates/radius (29.9N, 82.4W)/95.8 km, Population 375,761, Observed cases 530, Expected cases 420. The
yellow area represents the Central Florida cluster. SaTScanTM
computed results include: Coordinates/radius (28.2N, 81.5W)/13.4
km, Population 9,213, Observed cases 31, Expected cases 11.

compared with the state of Florida, those in this area are almost
three times as likely to be diagnosed with childhood cancer (Fig. 1).
Since a purely spatial analysis for the period 20002007 does not
indicate when the cluster appeared, a spacetime analysis was
performed, assessing these clusters using the Poisson model within
SaTScanTM. We observed that the spatial dimensions of the clusters
persisted during these periods. SF emerged as the most likely
temporal cluster with elevated risk during 20062007 (Fig. 2).
Whereas the observed cases were 403, the expected were 274,
RR 1.52, P 0.001, implying a significant 52% increase in
childhood cancer rate in SF compared with the state of Florida.
Similarly, the NEF emerged as a secondary temporal cluster for
20012004, with the observed and expected cases as 136 and 87
respectively, RR 1.59, P 0.06. This suggested a 59% increase in
the rate of overall childhood cancer in NEF relative to the state, but
the increase was not statistically significant.
To confirm the clusters, we compared cancer rates within SF to
the state. The cancer rates of the state for this time period was 14.1
per 100,000 in 2005 and increased slightly to 16.4 per 100,000 and
15.7 per 100,000 for 2006 and 2007, respectively. By contrast, from
2000 to 2007, the SF cancer rates have been consistently higher than
the corresponding Florida rates. In particular, the rates computed for
2006 and 2007 increased significantly from 13.8 per 100,000 in
2005 to 23.9 and 21.1 per 100,000, in 2006 and 2007, respectively.

514

Amin et al.
TABLE I. Childhood Age and Sex Adjusted Cancer Incidence
Rates for Florida, SF Cluster, and Florida Without SF Cluster
Area
FL

FL w/o SF

SF

Year

Rate

95% CI

Rate ratio

2006
2007
Aggregate
2006
2007
Aggregate
2006
2007
Aggregate

16.4
15.7
16.05
14.8
14.5
14.65
23.9
21.1
22.5

15.1, 17.6
14.5, 16.9
14.8, 17.3
13.5, 16.1
13.2, 15.8
13.2, 15.7
20.3, 27.5
17.7, 24.4
19.2, 26.1

1.0
1.0
1.0
0.9024
0.9236
0.9128
1.4573
1.3439
1.4019

Incidence counts were utilized directly to compute incidence rates using


the FAPTP Dataset for 20002007 and Florida population statistics for
2000. Confidence intervals are provided, as uncertainty still exists
within ideal registry datasets and computed cancer statistics (United
States Cancer Statistics: 1999 incidence). Aggregate refers to the rates
for 2006 and 2007 combined. The state of Florida is the reference group,
hence the ratio is 1.0 for FL, Florida. SF is the southern Florida cluster.
The time frame for the SF cluster was noted to be January 1, 2006 to
December 31, 2007. CI, Confidence intervals.
Fig. 2. Spacetime analysis of FAPTP database for all cancer types
20002007. Clustering representation of SaTScanTM spacetime
analysis is illustrated utilizing zip code data with age, sex, and race as
covariates. Clusters are represented in colors. Spatial representations
were not affected significantly however time frame results for the
Southern Florida (SF) cluster (20062007) are noted to be representative of a recent surge in incidence rates. The red area represents the
South Florida cluster. SaTScanTM computed results include: Coordinates/radius (26.0N, 81.4W)/121.1 km, Time frame January 1,
2006 to December 31, 2007, Population 963,643, Observed cases
403, Expected cases 274. The orange area represents the North
Central Florida cluster. SaTScanTM computed results include: Coordinates/radius (29.5N, 82.0W)/65.9 km, Time frame January 1,
2001 to December 31, 2004, Population 155,681, Observed cases
136, Expected cases 87.

However, when we excluded the SF cases from the overall Florida


cancer cases, the rates in Florida significantly decreased (Table I,
Fig. 3).
Purely spatial analysis of leukemia rates identified two regions of
Florida (during the period of 20002007) similar to the cluster
areas identified when all cancer types were combined. A total of
1,254 leukemia cases in the state were identified and utilized in
this analysis. There was a statistically significant cluster in SF
(RR 1.53, P 0.001) (Fig. 4). A second cluster was identified in
the north central region of the state, shifting somewhat from the NEF
cluster and was statistically significant as well, RR 1.45, P 0.03.
Likewise, in the spacetime analysis of leukemia cases, there was a
statistically significant cluster in SF (RR 1.74, P 0.05) (Fig. 5).
The time period identified for the peak rate of the cluster was 2000
2002. During this time period, the number of observed cancer cases
was 105 while the expected number of cases was 63. While the
spacetime analysis points to 20002002 as the time of the peak in
leukemia rates, the purely spatial analysis indicated that leukemia
rates in the SF cluster area remained elevated throughout the entire
period (20002007), when compared to the state.
A purely spatial analysis of brain/CNS cancer identified one area
in southern Florida. Of the 839 cases identified in the state, there
Pediatr Blood Cancer DOI 10.1002/pbc

were 60 observed and 33 expected cases in this region. The relative


risk comparing Florida to SF was not statistically significant,
RR 1.86, P 0.07 (Fig. 6). A spacetime analysis (52 observed
cases and 24 expected cases) for the brain/CNS cancer identified a
cluster corresponding to the SF cluster, with a statistically
significant increased incidence rate RR 2.25, P 0.02, implying
that children in SF were two times as likely to develop brain/CNS
cancer when compared with children in the state of Florida. The time
period identified for this cluster was 20062007 (Fig. 7). In contrast,
lymphoma rates were not statistically significant probably due to
small numbers.

Fig. 3. Age-adjusted pediatric cancer incidence rates 20002007.


Incidence counts were utilized directly to compute incidence rates using
FAPTP Dataset for 20002007 and Florida population statistics for
2000. Southern Florida cluster (SF) is shown in comparison to rates for
the entire state of Florida and to rates for the state of Florida excluding
the influence of the SF. Differences between these rates during 2006 and
2007 suggest that the rise in Florida rates during this period was
influenced by the surge in incidence rates in the SF cluster.

Florida Childhood Cancer Clusters

515

Fig. 4. Purely spatial analysis of leukemia cases 20002007.


Clustering representation of SaTScanTM purely spatial analysis is
illustrated utilizing FAPTP zip code data with age and sex covariates.
Clusters are represented in colors. The SF cluster remains durable and
significant with respect to the specific leukemia cases in Florida. The red
area represents the South Florida cluster. SaTScanTM computed results
include: Coordinates/radius (26.2N, 81.7W)/141.6 km, Population
417,327, Observed cases 190, Expected cases 131. The orange area
represents the North Central Florida cluster. SaTScanTM computed
results include: Coordinates/radius (29.1N, 82.7W)/120.1 km, Population 435,669, Observed cases 190, Expected cases 138.

Fig. 5. Spacetime analysis of leukemia cases 20002007. Clustering representation of SaTScanTM Spacetime analysis is illustrated
utilizing FAPTP zip code data with age and sex as covariates. Space
time clusters were statistically significant. Time frame results for the
Southern Florida (SF) cluster were noted to be 20002002. The yellow
area represents the South Florida cluster. SaTScanTM computed results
include: Coordinates/radius (26.0N, 81.6W)/128.5 km, Time frame
January 1, 2000 to December 31, 2002, Population 553,592, Observed
cases 105, Expected cases 63.

DISCUSSION

clusters, in the human population [2931,35,36]. By utilizing the


data from FAPTP, we ensured the accuracy and reliability of the data
used. FAPTP routinely reviews the cancer data for discrepancies
including duplications and provides the most comprehensive
incidence data of childhood cancer in Florida. Thus, FAPTP
facilitates assessment of patterns of cancer rates and geographical
trends within the state of Florida. Whereas the limitations addressed
in previous studies on clusters could not be avoided completely,
our chances of repeating similar methodologic issues were
substantially minimized as described below.
The large sample of cases with overall childhood cancer as well
as significant cases in cancer subtypes should ensure a sufficiently
high statistical power. It has been shown in a simple power study
[36] for the likelihood ratio test used in SaTScanTM that a relative
risk of 1.35 can result in an estimated power (1 " b > 0.80) to detect
the differences in cancer cases between the clusters and non-cluster
areas (in the state of Florida), if one does exist. For example, from
2006 to 2007, the observed cases were 403 in SF, which is a large
sample for comparison between areas with and without clusters
(Fig. 2). Because we used cancer data from a highly reliable source
(FAPTP), both selection and misclassification biases were
dramatically minimized in our study. The observed clusters in SF
and NEF are not driven by improved case ascertainment following
the increased childhood cancers in certain geographic areas in
Florida. In addition, because this study started 2 years ago, it is
highly unlikely that our findings are influenced by other recent
studies on Florida clusters.

These purely spatial and spacetime clustering studies of


childhood cancer in Florida were conducted using data from FAPTP
and the Census data of 2000. The accuracy of case ascertainment is
high with FAPTP and has been described and validated elsewhere
[28,32,33]. This epidemiologic mapping study of Florida reveals
three major findings. First, childhood cancer clusters were identified
in SF and NEF. Second, the childhood cancer clusters persisted
after controlling for age, sex, and race/ethnicity. Third, whereas
significant increase in cancer rates was observed in leukemia and
brain/CNS cancer, there was no significant increase in the
lymphoma rate among children in SF and NEF.
There are several methodologic issues in identification and
confirmation of childhood cancer clusters, especially leukemia [42].
In general, these studies are limited by low statistical power [43].
Therefore, the identification of cancer clusters may be driven by bias
such as the practice of defining geographical boundaries of the
cluster and improved case ascertainment in the areas suspected of
having clusters, as well as error, namely, random variation [44].
Cancer cluster studies utilizing multiple comparisons over a small
period of time or different methods have shown false positive results
[45]. Further, population density, age, migration, sex, and race/
ethnicity are potential confounding elements affecting childhood
cancer cluster confirmation [46,47].
This study utilized statistical software (SaTScanTM) that is
reliable in the assessment of cancer clusters, as well as other disease
Pediatr Blood Cancer DOI 10.1002/pbc

516

Amin et al.

Fig. 6. Purely spatial analysis of brain/CNS Tumor cases 20002007.


Clustering representation of SaTScanTM purely spatial analysis is
illustrated utilizing FAPTP zip code data with age and sex covariates.
The SF cluster significant although size of area is altered with respect
to prior cluster maps identified in Florida. The red area represents
the South Florida cluster. SaTScanTM computed results include:
Coordinates/radius (26.0N, 80.4 W)/15.6 km, Population 157,361,
Observed cases 60, Expected cases 33.

To better understand the increased cancer rates in SF, it is


important to consider changes in the population for that region as
well. Otherwise, the possible environmental factors affecting cancer
rates could be confounded with population migrations and
increases. While estimates for the pediatric population counts for
all ages in each of the Florida zip code areas were not available for
the period 20012007, we utilized population estimates for the
pediatric population by county for 2001 and 2007 from the Florida
Legislature [48] and the estimates of the pediatric population for a 3year-period 20052007 from the American Community Survey of
the Census Bureau [49]. Considering the relative annual population
increase, defined by the ratio r as follows:
r

pediatric pop 2007 " pediatric pop 2000


pediatric pop 2000

where the change in the pediatric population count in 2007 is


obtained relative to the pediatric population count in 2000, we
compared the average values for the ratio r for the SF area with the
corresponding annual relative population increase for the rest of
Florida. Similarly, we also obtained a ratio based on the 20052007
estimates. Our results indicated that relative population increases in
the SF cluster area are not significantly different from the rest of the
state. It is also possible that zip code population shifts over time
could have altered the results between 2000 and 2007. Such shifts
could result in an apparently elevated cancer rate when using 2000 as
the population standard. Using population estimates for larger areas
such as counties would limit the effects of such small-area
migrations on the cancer rates. Florida county population estimates
Pediatr Blood Cancer DOI 10.1002/pbc

Fig. 7. Spacetime analysis of brain/CNS Tumor Cases 20002007.


Clustering representation of SaTScanTM Spacetime analysis is
illustrated utilizing FAPTP zip code data with age and sex as covariates.
Clusters are represented in colors. The red area represents a Northeastern Florida cluster. SaTScanTM computed results include: Coordinates/
radius (30.1N, 81.8W)/20 km, Time frame January 1, 2005 to
December 31, 2007, Population 111,133, Observed cases 29,
Expected cases 9. The orange area represents the South Florida
cluster. SaTScanTM computed results include: Coordinates/radius (26.3N, 81.3W)/105.2 km, Time frame January 1, 2006 to
December 31, 2007, Population 455,519, Observed cases 52,
Expected cases 24.

between 2005 and 2007 were available for 53 counties in Florida


with populations greater than 20,000. We analyzed purely spatial
and spacetime SaTScanTM results for these 53 counties from 2000
to 2007 and found that the brain tumor cluster persisted. Analysis of
leukemia clusters persisted during the spacetime analysis but not
for the purely spatial analysis. While our initial analysis was based
on zip codes, limited analysis based on counties indirectly suggests
that population shifts did not play a significant role in altering the
cancer clusters. Thus, it is highly unlikely that our findings of
childhood cancer clusters are driven primarily by migration since
population changes in these geographic areas were non-differential,
thus minimizing any misclassification bias and confounding from
the observed clusters.
In this study, we have shown that there is a relative increase in
childhood cancer crude incidence rate in SF and NEF during
the years 20002007. Since this finding might have been influenced
by potential confounders of childhood cancer [44], we adjusted for
age at diagnosis, sex,and race/ethnicity and still observed a
statistically significant relative increase in SF and NEF compared
with the state of Florida. Therefore, given these adjustments, it is
possible to suspect geographic variation as the potential risk
variable for the clusters. Although the cluster areas identified are
quite large geographically, it is possible that localized environmental factors or person-to-person spread of viral or bacterial
pathogen [12,13,2124], may be involved in these suspected

Florida Childhood Cancer Clusters


geographic areas. Finally, despite these adjustments, we cannot rule
out unmeasured confounding elements as a possible explanation of
the observed clusters. Furthermore, residual confounding elements
may influence this confirmation especially by race/ethnicity, since
this information may have suffered from misclassification bias.
Therefore, statistical modeling cannot completely remove the effect
of confounding [27,50].
Our study found the crude incidence rate of childhood leukemia
and brain/CNS cancers to be significantly higher in the SF and NEF
clusters when compared with the state of Florida. As described
earlier, these findings are unlikely to be driven by non-factual
attributes of cancer clusters but are suggestive of environmental
factors or common risk factors in the areas. Consequently, these
findings could be etiologically driven, indicating the need for further
investigation to identify the potential risk factors in the observed
leukemia and brain/CNS cancer clusters in these areas. We did not
find spatial or spacetime clustering with lymphoma in the adjusted
models. The negative finding with lymphoma may be due to the
small number of cases in this subset, which limits the statistical
power to detect significant clusters with these data [47] or due to the
lack of a lymphoma cluster.
Despite the strengths of this article, there are also some
limitations. First, we used a preexisting dataset that may be
associated with information and selection bias, thus influencing the
validity of our findings. However, since the FAPTP data are highly
reliable, it is unlikely that our confirmation of cancer clusters in SF is
driven solely by information or selection bias. Second, confounding
elements such as race/ethnicity and age may very well influence our
results. But this is unlikely since we focused on childhood
malignancy with no reference to adult tumors. Finally, as with all
epidemiologic studies, unmeasured and residual confounding
elements may also partly influence the findings reported.
In summary, we found evidence of spatial and spacetime
childhood cancer clustering in SF and NEF. Statistically significant
cancer subtype clustering was found for leukemia and brain/CNS
cancer but not for lymphomas, which may be due to low statistical
power of our study to detect smaller clusters. This evidence is
suggestive of the presence of some environmental and possibly
social conditions that may act individually or collectively to
predispose children in these cluster regions to increased risk of
childhood cancer. Further study is needed to investigate the possible
predisposing factors in the elevated childhood cancer rates in SF and
NEF.

ACKNOWLEDGMENT
This study was supported in part by Nemours Childrens Clinic,
Pensacola, the Nemours Foundation and the Caitlin Robb
Foundation. We also thank Gulf Coast Wings of Hope for their
support. We thank Brian Calkins and Wendy McLeod of FAPTP for
their assistance in obtaining data and Dr. Elliot Daniel, Christiana
Care, DE, for the critical reading of the manuscript.

REFERENCES
1. Heron MP, Hoyert DL, Murphy SL, et al. Deaths: Final data for
2006. National vital statistics reports. Hyattsville, MD: National
Center for Health Statistics; 2009. http://www.cdc.gov/nchs/
fastats/children.htm. 57(14) Available from URL.
Pediatr Blood Cancer DOI 10.1002/pbc

517

2. American Cancer Society. Cancer Facts and Figures


2007. Available from http://www.cancer.org/downloads/STT/
CAFF2007PWSecured.pdf.
3. American Cancer Society. Cancer in Children. http://www.cancer.
org/docroot/CRI/content/CRI_2_4_1X_Introduction_7.asp?rnav
cri (accessed 9/1/2009).
4. Strahm B, Malkin D. Hereditary cancer predisposition in children:
Genetic basis and clinical implications. Int J Cancer 2006;119:
20012006.
5. Pakakasama S, Tomlinson GE. Genetic predispositions and
screening in pediatric cancer. Pediatr Clin North Am 2002;49:
13931413.
6. Davies SM. Subsequent malignant neoplasms in survivors of
childhood cancer: Childhood cancer survivor study (CCSS)
studies. Pediatr Blood Cancer 2007;48:727730.
7. Shuryak I, Hahnfelt P, Hlatky L, et al. A new view of radiation
induced cancer: Integrating short- and long-term processes. Part II:
Second cancer risk estimation. Radiat Environ Biophys 2009;48:
275286.
8. Scott D. Chromosomal radiosensitivity and low penetrance predisposition to cancer. Cytogenet Genome Res 2004;104:365370.
9. Taylor M, Hussain A, Urayama K, et al. The human major
histocompatibility complex and childhood leukemia: An etiological hypothesis based on molecular mimicry. Blood Cells Mol Dis
2009;42:129135.
10. Pereira TV, Rudnicki M, Pereira AC, et al. 5,10-Methylenetetrahydrofolate reductase polymorphisms and acute lymphoblastic
leukemia risk: A meta-analysis. Cancer Epidemiol Biomarkers
Prev 2006;15:19561963.
11. Ogilvy-Stuart AL, Gleeson H. Cancer risk following growth
hormone use in childhood: Implications for current practice. Drug
Saf 2004;27:369382.
12. McNally RJ, Eden TO. An infectious aetiology for childhood acute
leukemia: A review of the evidence. Br J Haematol 2004;127:243
263.
13. Buffler PA, Kwan ML, Reynolds P, et al. Environmental and
genetic risk factors for childhood leukemia: Appraising the
evidence. Cancer Invest 2005;23:6075.
14. Kim AS, Eastmond DA, Preston RJ. Childhood acute lymphocytic
leukemia and perspectives on risk assessment of early-life stage
exposures. Mutat Res 2006;613:138160.
15. Chang JS. Parental smoking and childhood leukemia. Methods Mol
Biol 2009;472:103137.
16. Dal Maso L, Bosetti C, LaVecchia C, et al. Risk for thyroid cancer:
An epidemiological review focused on nutritional factors. Cancer
Causes Control 2009;20:7586.
17. Infante-Rivard C, Weichenthal S. Pesticides and childhood cancer:
An update of Zahm and Wards 1998 review. J Toxicol Environ
Health B Crit Rev 2007;10:8199.
18. Howard SC, Metzger ML, Wiliams JA, et al. Childhood cancer
epidemiology in low-income countries. Cancer 2008;112:461
472.
19. Wigle DT, Arbuckle TE, Turner MC, et al. Epidemiologic evidence
of relationships between reproductive and child health outcomes
and environmental chemical contaminants. J Toxicol Environ
Health B Crit Rev 2008;11:373517.
20. Schieve LA, Rasmussen SA, Buck GM, et al. Are children born
after assisted reproductive technology at increased risk for adverse
health outcomes? Obstet Gynecol 2004;103:11541163.
21. Bellec S, Hemon D, Rudant J, et al. Spatial and spacetime
clustering of childhood acute leukaemia in France from 1990 to
2000: A nationwide study. Br J Cancer 2006;94:763770.
22. McNally RJ, Alexander FE, Bithell JF. Spacetime clustering of
childhood cancer in Great Britain: a national study, 19691993. Int
J Cancer 2006;118:28402846.

518

Amin et al.

23. McNally RJ, Bithell JF, Vincent TJ, et al. Spacetime clustering of
childhood cancer around the residence at birth. Int J Cancer
2009;124:449455.
24. McNally RJ, Alexander FE, Vincent TJ, et al. Spatial clustering of
childhood cancer in Great Britain during the period 19691993. Int
J Cancer 2009;124:932936.
25. Lawson AB. Statistical Methods in Spatial Epidemiology. 2nd
edition. New York: John Wiley & Sons; 2006.
26. Kulldorff M. Information Management Services, Inc. SaTScanTM
v8.0: Software for the Spatial and SpaceTime Scan Statistics,
2009. http://www.SaTScan.org/.
27. Thun MJ, Sinks T. Understanding cancer clusters. CA Cancer J Clin
2004;54:273280.
28. Pinheiro PS, Button JH, Fleming LE, et al. Pediatric cancer in
Florida, 19812000. Miami, FL: Florida Cancer Data System;
Available from URL: http://fcds.med.miami.edu/inc/statistics.
shtml.
29. Kulldorff M, Feuer EJ, Miller BA, et al. Breast cancer clusters in the
northeast United States: A geographic analysis. Am J Epidemiol
1997;146:161170.
30. Kulldorff M, Athas WF, Feurer EJ, et al. Evaluating cluster
alarms: A spacetime scan statistic and brain cancer in Los
Alamos. New Mexico: Am J Public Health; 1988. pp. 1377
1380.
31. Kearney G. A procedure for detecting childhood cancer clusters
near hazardous waste sites in Florida. J Environ Health 2008;70:
2934.
32. Roush SW, Krischer JP, Cox MW, et al. Progress in Childhood
Cancer Care in Florida 19701992. J Fla Med Association
1993;80:747751.
33. Krischer JP, Roush SW, Cox MW, et al. Using a population-based
registry to identify patterns of care in childhood cancer in Florida.
Cancer 1993;71:33313336.
34. Steliarova-Foucher E, Stiller C, Lacour B, et al. International
Classification of Childhood Cancer, third edition. Cancer 2005;
103:14571467.
35. Knox E. Detection of clusters. In: Elliott P, editor. Methodology of
enquiries into disease clustering. Small area health statistics unit.
1989.

Pediatr Blood Cancer DOI 10.1002/pbc

36. Kulldorff M, Nagarwalla N. Spatial disease clusters: Detection and


inference. Stat Med 1995;14:799810.
37. Kulldorff M. A spatial scan statistic. Comm Stat Theory Methods
1997;26:14811496.
38. Woods & Poole Economic, Inc., 1794; Columbia Road, N.W. Suite
4 Washington, DC 20009-2808 . http://woodsandpoole.com/.
39. Hansen KM. Head-banging: Robust smoothing in the plane. EE
Trans Geosci Remote Sens 1991;29:369378.
40. Hansen KM. and Statistical Research and Applications Branch,
NCI (2003). Headbang Software v3.0. http://srab.cancer.gov/
headbang/.
41. Mungiole M, Pickle LW, Simonson KH. Application of a weighted
head-banging algorithm to mortality data maps. Stat Med 1999;18:
32013209.
42. Hyams J. Leukemia and the clusters enigma: The towns united to
fear. Australian Women Weekly 1988;Oct: 44-6.
43. Linet MS. The Leukemia: Epidemiologic aspects. New York:
Oxford University Press; 1985.
44. Fraumeni JF, Jr., Miller RW. Epidemiology of human leukemia:
Recent observations. J Natl Cancer Inst 1967;38:593605.
45. Wartenbeberg D, Greenberg MM. Detecting disease clusters: The
importance of statistical power. Am J Epidemiol 1990;132:S156
S166.
46. Wartenbeberg D, Greenberg MM. Solving the cluster puzzle: Clues
to follow and pitfalls to avoid. Statist Med 1993;12:17631770.
47. Alexander FE, Boyle P, In: editors. Methods for investigating localized clustering of disease. IARC scientific publication
No 135. Lyon: International Agency for Research on Cancer; 1996.
pp. 1247.
48. The Office of Economic and Demographic Research (EDR),
Florida Legislature, http://edr.state.fl.us/conferences/population/
demographic.htm.
49. U.S. Census Bureau, http://factfinder.census.gov/servlet/DatasetMainPageServlet?_programACS&_submenuId&_langen&_
ts.
50. Holmes L, Chan W, Jiang Z, et al. Effectiveness of androgen
deprivation therapy in prolonging survival of men treated for locoregional prostate cancer. Prostate cancer and Prostatic Dis 2007;10:
388395.