Ecoepidemiology of Tularemia in the Southcentral United States
Rebecca J. Eisen,* Paul S. Mead, Andrew M. Meyer, Liza E. Pfaff, Kristy K. Bradley, and Lars Eisen
Division of Vector-Borne Infectious Diseases, National Center for Zoonotic, Vector-Borne, and Enteric Diseases,
Centers for Disease Control and Prevention, Fort Collins, Colorado; Department of Microbiology, Immunology, and Pathology,
Colorado State University, Fort Collins, Colorado; Office of the State Epidemiologist, Oklahoma State Department of Health,
Oklahoma City, Oklahoma
kansas, Illinois, Indiana, Kansas, Kentucky, Missouri, Nebraska, Oklahoma, and Tennessee) in the southcentral United
States with Geographic Information System (GIS)-based environmental data to determine associations between cov-
erage by different habitats (especially dry forest representing suitable tick habitat) and tularemia incidence. High-risk
counties (> 1 case per 100,000 person-years) clustered in Arkansas-Missouri and far eastern Oklahoma and Kansas.
County tularemia incidence was positively associated with coverage by dry forested habitat suitable for vector ticks for
Oklahoma-Kansas-Nebraska and Arkansas-Missouri but not for Illinois-Indiana-Kentucky-Tennessee. A multivariate
logistic regression model predicting presence of areas with risk of tularemia based on GIS-derived environmental data
was developed for the Arkansas-Missouri tularemia focus. The study shows the potential for research on tularemia
ecoepidemiology and highlights the need for further modeling efforts based on acarologic data and more fine-scale point
or zip code/census tract epidemiologic data.
We combined county-based data for tularemia incidence from 1990 to 2003 for a nine-state region (Ar-
Tularemia is caused by the bacterium Francisella tularensis,
which can be transmitted to humans through a variety of
routes including tick or insect bite, handling of infected ani-
mals, contact with or ingestion of infected water, food, or soil,
and inhalation of infectious aerosol.1–4Approximately 90%
of North American tularemia cases are considered to be
caused by F. tularensis tularensis (Type A),5but tularemia
caused by F. tularensis holarctica (Type B) may be underdi-
agnosed or underreported. Several aspects of the epidemiol-
ogy of tularemia in the United States have changed dramati-
cally since the early 1900s. First, the average of 123 cases
reported annually during 1990–2004 is an order of magnitude
lower than the average of 1,181 cases reported annually dur-
ing 1940–1948.2,6–8This tremendous decline in national tula-
remia cases, which occurred primarily during the 1950s and
1960s,9most likely was related to a general decrease in human
exposure to infected animals, particularly rabbits. Second, the
spatial distribution of tularemia cases changed dramatically
over the 20th century. The current national focus of tularemia
in Arkansas-Missouri7is distinctly smaller than the historical
one, with large numbers of cases occurring also in other east-
ern states (e.g., Georgia, Illinois, Indiana, Kentucky, Louisi-
ana, Mississippi, Ohio, Tennessee, and Virginia).2It was re-
cently argued that ticks historically and currently play a more
prominent role in the transmission of F. tularensis to humans
in Arkansas-Missouri than in other parts of the eastern
United States.10This may be the primary reason for a ∼2-fold
reduction in tularemia in this two-state area, whereas inci-
dence decreased ∼8-fold from other eastern states after the
decline in human exposure to infected lagomorphs.10This
notion is supported by the exceptional numbers of vector
ticks (Dermacentor variabilis and, especially, Amblyomma
americanum) that can infest dry forested habitats in Arkan-
The epidemiology of tularemia in the southcentral and
southeastern United States was studied intensively from the
late 1940s to the late 1980s13–25but has received virtually no
attention over the last two decades. Recent advances in Geo-
graphic Information System (GIS) technology and increasing
availability of GIS-based environmental data have opened
new avenues for ecoepidemiologic studies of associations be-
tween environmental factors and tularemia incidence and the
intriguing ecology of the tularemia agent. This study, which
focused on a nine-state region in the southcentral United
States (Arkansas, Illinois, Indiana, Kansas, Kentucky, Mis-
souri, Oklahoma, Nebraska, and Tennessee), represents an
initial effort to use GIS technology to explore the linkages
between the environment and risk of tularemia in the United
States. Primary aims were to 1) determine associations be-
tween coverage by different habitat types (especially dry for-
est representing suitable tick habitat) and tularemia incidence
in the nine-state region and 2) develop a multivariate logistic
regression model predicting presence of areas with risk of
tularemia in the Arkansas-Missouri disease focus based on
associations between GIS-derived environmental data and tu-
MATERIALS AND METHODS
Epidemiologic data. In the United States, tularemia is a
nationally notifiable disease. Health care providers are re-
quired by law to report cases to state or local health officials,
who in turn report cases to the Centers for Disease Control
and Prevention (CDC) through the National Notifiable Dis-
eases Surveillance System (NNDSS). Both confirmed and
probable cases are tallied at the national level. A confirmed
case is defined as clinically compatible illness with isolation of
F. tularensis from a clinical specimen or a 4-fold or greater
change in antibody titer to F. tularensis antigen.26A probable
case is defined as clinically compatible illness with elevated
serum antibody titer(s) to F. tularensis antigen (without a
documented 4-fold change in titer) or detection of F. tularen-
sis in a clinical specimen by fluorescent assay.26More than
80% of reported cases meet the confirmed definition. Al-
though reporting of tularemia cases was officially discontin-
ued from 1995 through 1999,27cases continued to be reported
* Address correspondence to Rebecca J. Eisen, PO Box 2087, Fort
Collins, CO 80522. E-mail: firstname.lastname@example.org
Am. J. Trop. Med. Hyg., 78(4), 2008, pp. 586–594
Copyright © 2008 by The American Society of Tropical Medicine and Hygiene
to the CDC at a steady rate during this period (CDC, unpub-
lished data). Underreporting of tularemia cases has not been
systematically studied but is assumed to occur. However, con-
sistent underreporting should not interfere with comparisons
over time or across jurisdictions.
Our study included data for county-based tularemia inci-
dence from 1990 to 2003 for a nine-state region of the south-
central United States: Arkansas, Illinois, Indiana, Kansas,
Kentucky, Missouri, Oklahoma, Nebraska, and Tennessee.
These states accounted for 838 (72%) of 1,169 confirmed and
probable cases reported to CDC over this 14-year period
(CDC, unpublished data). County incidences were calculated
based on human population census data from 2000 (ESRI,
Redlands, CA). The study region was subdivided into three
state groupings based on similarities in general climate and
land cover types: Oklahoma-Kansas-Nebraska to the west
(with large tracts of prairie), Arkansas-Missouri in the center
(with large tracts of coniferous and mixed deciduous/
coniferous forest in northern Arkansas and southern Mis-
souri), and Illinois-Indiana-Kentucky-Tennessee to the east
(with agricultural landscapes interspersed with predominantly
deciduous forest) (Table 1). Because of vast ecological differ-
ences within the targeted nine-state area, analyses were run
separately for the aforementioned ecologically similar state
groupings. Within each of these state groupings, county-based
incidence of tularemia ranged from 0 to > 1 case per 100,000
person-years, with the majority of counties reporting no cases
from all states expect Arkansas and Missouri (Figure 1). We
also evaluated the possibility of incorporating other adjoining
southcentral, southeastern, or northcentral states into the
study (Texas, Mississippi, Louisiana, Alabama, Georgia,
Ohio), but these were ultimately excluded because tularemia
incidences were very low (mean county tularemia incidence
by state of < 0.01 case per 100,000 person-years), and varia-
tion in incidence among counties was insufficient for mean-
ingful analysis at this spatial scale.
Environmental data. GIS-based data used in the study in-
cluded 1) administrative boundaries (state, county; ESRI); 2)
elevation (30 × 30-m spatial resolution; US Geological Survey
national digital elevation data set); 3) long-term (1961–1990)
average climate data (mean annual, January, and July mini-
mum, mean, and maximum temperature; mean annual cool-
ing and heating degree days and growing degree-days; mean
annual, January, and July precipitation and relative humidity;
mean annual snowfall; median annual length of freeze-free
period; median Julian date of first and last snowfall; 2 × 2-km
spatial resolution; Climate Source, Corvallis, OR); 4) annual
average normalized difference vegetation index (NDVI) data
from 2005 (1 × 1-km spatial resolution; derived from NOAA
Advanced Very High Resolution Radiometer images); and 5)
land cover classifications based on US Geological Survey
state Gap analysis projects. Land cover was reclassified from
state-specific classification schemes into a uniform classifica-
tion scheme including the following habitat types: water, bar-
ren, urban, agricultural crops, seasonally flooded habitat, wet-
land, grassland (including pasture), shrubland, dry deciduous
forest, dry coniferous forest, and dry mixed forest. These
habitat types fall into three broad categories with regards to
suitability for tick vectors: unsuitable (water, barren, agricul-
tural crops, seasonally flooded habitat, wetland), partially
suitable (grassland, shrubland, urban), and suitable (dry de-
ciduous forest, dry coniferous forest, dry mixed forest). For
the purpose of this study, dry forest refers to forested habitats
not prone to seasonal flooding. The rationale for including
grassland in the partially suitable, rather than suitable, cat-
egory was that the human-biting life stages of both the Ameri-
can dog tick (D. variabilis) and the lone star tick (A. ameri-
canum) are abundant in dry forested habitats, but that only D.
variabilis can be considered potentially abundant in grass-
lands.12,28,29Furthermore, grasslands likely are only partially
suitable even for D. variabilis in the dry environments char-
acteristic of the western part of the targeted nine-state area.30
It is important to note that the broad habitat classifications of
suitable, partially suitable, and unsuitable for vector ticks
were developed to elucidate associations with tularemia inci-
dence at the crude county scale; they do not account for more
fine-scale landscape patterns potentially related to tick abun-
dance such as presence of habitat ecotones (e.g., forest-
grassland edges or forested borders of the peridomestic en-
vironment) or presence of small patches of a given habitat
within another one (e.g., very small patches of dry forest
within an area characterized by seasonal flooding). Further-
more, old field habitats, which are known to present risk for
exposure to D. variabilis,29,30likely were separated in the
Gap land cover classification into grassland, shrubland, or
forest categories depending on the degree of reforestation of
an individual field at the time the classification was created.
Determination of associations between habitat type and
tularemia incidence. County unit-based associations between
the above-mentioned habitat types (percentages of coverage)
and 12 tularemia incidence classes (0, 0.001–0.049, 0.050–
0.099, 0.100–0.199, 0.200–0.299, 0.300–0.399, 0.400–0.499,
0.500–0.749, 0.750–0.999, 1.000–1.499, 1.500–1.999, and 2.000–
6.500 cases per 100,000 person-years) were tested by ordinal
logistic regression. Incidence classes were chosen to have
similar numbers of counties (range: 17–39) for each incidence
group where cases occurred. Ordinal logistic regression based
on tularemia incidence classes, rather than linear regression
based on tularemia incidences, was used because tularemia
incidence data were not normally distributed and transforma-
tion could not make them so. Separate tests were carried out
for each of the three state groupings included in the study:
Oklahoma-Kansas-Nebraska, Arkansas-Missouri, and Illi-
Development of predictive models for risk of exposure to
tularemia in Arkansas-Missouri. Because the vast majority of
cases were reported from Missouri and Arkansas and this
region differs ecologically from the other groupings (Figures
1 and 3), attempts to fit a model to the entire nine state region
obscured small-scale (i.e., county and sub-county) differences
within the Missouri-Arkansas tularemia focus. Similarly,
Yabsley and others31showed that modeling another tick-
borne disease, human mononcytic ehrlichiosis, in a similar
geographic region based on subregions yielded a better over-
all fit than the region-wide model. Within Arkansas-Missouri,
continuous environmental data (mean elevation, annual
NDVI, and climate variables as listed above) were used in a
multivariate logistic regression modeling approach aimed at
predicting the spatial pattern of presence of counties with
tularemia cases. Similar models were not developed for
Oklahoma-Kansas-Nebraska and Illinois-Indiana-Kentucky-
Tennessee because the numbers of tularemia cases for these
state groupings were judged too low to show associations at
the relatively crude county-based spatial scale.
Annual average NDVI, which is a measure of vegetation
based on visible and near-infrared light that is reflected by
vegetation, was found to serve as a proxy for dry forested
habitat suitable for ticks in the Arkansas-Missouri area; there
was a significant correlation between mean NDVI for a
county and the percentage of the county covered by dry for-
ested habitat (Spearman coefficient of rank correlation; ?s?
0.856, N ? 190, P < 0.001). Exclusive use of continuous co-
variates with a maximum spatial resolution of 2 × 2 km al-
lowed us to generate predictions for tularemia risk at a finer
spatial scale than the original county unit scale. The validity of
these fine-scale predictions will be examined in future studies.
Model development was based on a random selection
(within the 12 tularemia incidence classes mentioned previ-
ously) of 75% of the total counties in Arkansas-Missouri, with
the remaining 25% of counties set aside for model validation.
Models were developed for the probability that an area (i.e.,
a county polygon or a 2 × 2-km cell within a particular county)
would be classified as suitable for a tularemia case to occur.
Covariates included in the forward stepwise regression model
(probability to enter of 0.25) were restricted to variables
yielding significant associations with presence of tularemia in
univariate tests (Wilcoxon rank sum test with normal approxi-
mation; P < 0.05) and not strongly correlated with each other
(Spearman rank correlation; ?s< 0.8). The output model was
highly significant (whole model test; P < 0.005), and the lack
of fit test indicated that the model included sufficient num-
bers of covariates (P > 0.05). Covariates included in the
model are shown in Table 3.
The model is described by the following equation:
Logit ?P? = ?0+ ?1x1+ ?2x2+ ? ? ? + ?kxk
where P is the probability of a cell or a county polygon being
classified as a risk area and ?0is the intercept. The values
?1, . . . ,?krepresent the coefficients assigned to each indepen-
dent variable included in the regression, and x1, . . . ,xksym-
bolize the independent variables. The probability that a par-
ticular cell or county polygon in the GIS is classified as risk or
high risk, depending on the model, of exposure to the tulare-
mia agent can be derived from equation 1 using the following
P = exp??0+ ?1x1+ ? ? ? + ?kxk??
?1 + exp??0+ ?1x1+ ? ? ? + ?kxk??
At the county scale, each county was assigned a single
probability value. The optimal probability cut-off value was
chosen by maximizing sensitivity and specificity simulta-
neously using receiver operating characteristic (ROC) curves.
Counties with values above the probability threshold value
were classified as risk, whereas all others were considered no
At the sub-county scale for each county, the maximum
probability value among 2 × 2-km raster cells populating each
county was extracted using zonal statistics (ArcGIS9.2 Spatial
Analyst; ESRI). The optimal probability cut-off value was
chosen by maximizing sensitivity and specificity simulta-
neously. Counties containing at least one raster cell with a
probability value at least equal to the optimal value were
classified as containing risk areas, whereas all other counties
were considered to lack risk areas in our evaluation matrix.
Statistical analyses were carried out using Version 5.1 of the
JMP statistical package,32and the results were considered
significant when P < 0.05.
Spatial patterns of tularemia incidence. Mean county tula-
remia incidence per 100,000 person-years ranged from 0.02
for Indiana to 0.98 for Arkansas, and peak county values
ranged from 0.35 for Indiana to 6.21 for Missouri (Table 1).
Pairwise tests revealed that the Arkansas-Missouri grouping
had a higher county tularemia incidence per 100,000 person-
years (median, 0.52; mean, 0.80 ± 0.97) than either the Okla-
homa-Kansas-Nebraska grouping (median, 0; mean, 0.15 ±
0.44) or the Illinois-Indiana-Kentucky-Tennessee grouping
(median, 0; mean, 0.05 ± 0.14; Wilcoxon rank sum test with
normal approximation; z ? 11.20; df ? 1,463; P < 0.001 in
both cases). Although there was a trend toward higher county
tularemia incidence for the Oklahoma-Kansas-Nebraska
grouping than the Illinois-Indiana-Kentucky-Tennessee
grouping, the difference was not statistically significant (P ?
0.06). Counties with reported tularemia cases and, especially,
with tularemia incidences exceeding 1 case per 100,000 per-
son-years from 1990 to 2003 clustered in an area encompass-
ing Arkansas, southern Missouri, and eastern Oklahoma and
Kansas (Figure 1). We also were able to compare county
tularemia incidence for Arkansas during 1978–198221and
1990–2003 (Figure 2). This comparison showed both a general
County-based tularemia incidence and environmental characteristics for states included in the study
Mean ± SD (peak)
county incidence of tularemia
per 100,000 person-years,
Mean ± SD
of dry forest areas
Mean ± SD county climate data*
0.25 ± 0.59 (4.07)
0.15 ± 0.42 (2.99)
0.09 ± 0.28 (1.52)
20 ± 20
5 ± 8
3 ± 3
27.6 ± 0.6
26.0 ± 0.7
23.8 ± 0.9
949 ± 204
754 ± 200
633 ± 118
59.4 ± 3.3
59.8 ± 3.8
61.8 ± 3.4
172 ± 84
426 ± 106
747 ± 144 93
0.98 ± 0.95 (4.31)
0.68 ± 0.96 (6.21)
41 ± 26
32 ± 24
26.9 ± 0.8
25.3 ± 0.6
1,293 ± 80
1,056 ± 90
68.7 ± 2.2
67.5 ± 1.7
110 ± 86
419 ± 107
0.09 ± 0.19 (1.03)
0.02 ± 0.06 (0.35)
0.05 ± 0.15 (0.92)
0.04 ± 0.10 (0.41)
13 ± 10
21 ± 20
42 ± 24
47 ± 22
24.3 ± 1.0
23.6 ± 0.8
24.5 ± 0.9
24.9 ± 1.2
1,009 ± 88
1,053 ± 84
1,222 ± 74
1,380 ± 94
69.8 ± 1.4
69.4 ± 1.1
71.2 ± 1.4
72.5 ± 1.7
547 ± 171
583 ± 236
354 ± 64
228 ± 132
* Mean values for 1961–1990 based on GIS-derived data.
EISEN AND OTHERS
decline in tularemia incidence and a shift in the primary risk
area toward the north in Arkansas from 1978–1982 to 1990–
Determination of associations between habitat type and
tularemia incidence. The distribution of land cover types with
different perceived suitability for tick vectors of F. tularensis
(suitable: dry forested habitats; partially suitable: grassland,
shrubland, urban; unsuitable: water, barren, agricultural
crops, seasonally flooded habitat, wetland) in the nine-state
region is shown in Figure 3. County tularemia incidence was
strongly positively associated with coverage by dry forested
habitat suitable for vector ticks for the Arkansas-Missouri
and Oklahoma-Kansas-Nebraska groupings (P < 0.001 in
both cases; Table 2). A breakdown of dry forest categories
(deciduous, coniferous, mixed) showed positive associations
between tularemia incidence and all three forest categories
for both state groupings (P < 0.05 in all cases; Table 2). In
striking contrast, coverage by dry forested habitat suitable for
vector ticks was negatively associated with county tularemia
incidence for the Illinois-Indiana-Kentucky-Tennessee group-
ing. Other positive associations with county tularemia inci-
dence (P < 0.05) included coverage by grassland for the
Arkansas-Missouri and Illinois-Indiana-Kentucky-Tennessee
groupings and by water for the Oklahoma-Kansas-Nebraska
grouping (Table 2). Finally, negative associations with county
tularemia incidence (P < 0.05) included coverage by crop-
lands for the Arkansas-Missouri and Oklahoma-Kansas-
Nebraska groupings, by seasonally flooded habitats and wet-
lands for the Oklahoma-Kansas-Nebraska grouping, and by
shrubland or urban areas for the Illinois-Indiana-Kentucky-
Tennessee grouping (Table 2).
Predictive model for areas in Arkansas-Missouri with risk
of exposure to tularemia. A multivariate logistic regression
model was developed for presence of areas with risk of tula-
remia (Figures 4 and 5; based on annual NDVI, annual maxi-
mum temperature, annual relative humidity, and precipita-
tion in July; whole model test: ?2? 21.88, df ? 4, P < 0.001).
Lack of fit test indicated that the model included sufficient
numbers of covariates (P ? 0.35). Details for covariates in-
cluded in the model are shown in Table 3.
Before extrapolating the model to the sub-county scale, we
evaluated model performance at the county scale (Table 4;
Figure 4). Producer accuracy for classifying counties that re-
ported cases as counties at risk was 77.1% for the build set
and 70.6% for the validation set. For counties predicted to
pose a risk, 86.2% of the build set and 85.7% of the validation
set reported cases.
Next, we evaluated model performance at the sub-county
scale. Among the 105 counties in the build set that reported
tularemia cases, 89 (85%) were classified by our model as
containing at least one 2 × 2-km risk area. Similarly, 82% of
counties classified as containing risk areas reported tularemia
cases (Table 5; Figure 5). Among counties from which tula-
remia cases were not reported, 47% were classified as not
containing risk areas. For counties classified as not containing
incidence in Arkansas during 1978–198221and 1990–2003.
Comparison of spatial patterns of county tularemia
southcentral United States (Arkansas, Illinois, Indiana, Kansas, Ken-
tucky, Missouri, Oklahoma, Nebraska, and Tennessee) reporting tu-
laremia cases or with tularemia incidences exceeding 1 case per
100,000 person-years from 1990 to 2003. This figure appears in color
Distribution of counties in a nine-state region of the
ceived suitability for tick vectors of F. tularensis in the southcentral
United States (Arkansas, Illinois, Indiana, Kansas, Kentucky, Mis-
souri, Oklahoma, Nebraska, and Tennessee). Grass/shrub includes
prairie, pasture, and shrubland. Other habitats include water, barren,
agricultural crops, seasonally flooded habitat, and wetland. This fig-
ure appears in color at www.ajtmh.org.
Distribution of land cover types with different per-
risk areas, 53% did not report cases. Evaluation of the same
model based on counties that were not included in model
construction revealed similar accuracy (Table 5). For counties
reporting tularemia cases, 88% were classified by the model
as containing risk areas. When a county was classified as con-
taining risk areas, 88% of counties reported cases. Further-
more, 69% of counties not reporting cases were classified as
lacking risk areas and 69% of counties classified as lacking
risk areas did not report cases.
We present here the first GIS-based predictive spatial
model for tularemia risk in the United States. The study in-
cluded the national Arkansas-Missouri tularemia focus and
surrounding areas to the west (Oklahoma-Kansas-Nebraska)
and east (Illinois-Indiana-Kentucky-Tennessee). GIS-based
environmental data were found to be highly useful for studies
of the ecoepidemiology of tularemia, including the develop-
ment of a predictive spatial risk model for the Arkansas-
Missouri tularemia focus. Our modeling efforts both gener-
ated testable hypotheses for fine-scale patterns of tularemia
risk within Arkansas-Missouri and yielded new knowledge
regarding habitat associations of tularemia in a nine-state re-
gion that will help unravel the natural maintenance cycles of
the tularemia agent F. tularensis. This initial study also high-
lighted the need for further modeling efforts including a com-
bination of field-based data for acarologic risk of vector ex-
posure and more fine-scale point, zip code, or census tract–
based epidemiologic data.
As expected from previous studies including the southcen-
tral United States,7,23we found that counties with elevated
incidence of tularemia clustered in an area encompassing
western and northern Arkansas, southern Missouri, and far
eastern Oklahoma and Kansas (Figure 1). The distinct spatial
pattern of tularemia risk in Oklahoma and Kansas, with in-
creasing likelihood of counties with tularemia cases in the far
eastern parts of these states, is readily explained by land cover
type. Tularemia risk is low in the prairie landscapes to the
west but increases toward the Arkansas and Missouri borders
as the prairie is replaced by dry forested habitats suitable for
tick vectors (D. variabilis adults and A. americanum nymphs
and adults) of F. tularensis2,4,10,12(Figure 3). Although the
black-legged tick (Ixodes scapularis) has been implicated as a
vector of F. tularensis33and found to be naturally infected
with the tularemia agent in Oklahoma,33this tick can be con-
sidered to play a minor role in the transmission of F. tularensis
to humans in the southcentral United States because only the
adult stage commonly infests humans in this part of the coun-
try, and the adults are not active during the summer months
when most tularemia cases occur.10The positive association
between county tularemia incidence and coverage by water in
Oklahoma-Kansas-Nebraska is not surprising because ticks in
the dry prairie landscapes characteristic of these states un-
doubtedly cluster along water sources such as rivers and
The spatial patterns of counties with elevated incidence of
tularemia in Arkansas and Missouri (Figure 1) also can be
explained by the general distribution of dry forested habitat
suitable for vector ticks (Figure 3), which occurs widely in
western and northern Arkansas and southern Missouri but is
sparse in the Mississippi River valley of southeastern Arkan-
sas and the grasslands of northern Missouri where tularemia
in humans occurs only infrequently. Indeed, our predictive
spatial model for risk of tularemia in Arkansas-Missouri was
based in part on a variable indicative of dry forests (NDVI).
The association between habitat suitable for vector ticks and
tularemia incidence in Arkansas-Missouri is in accordance
with the most recent epidemiologic study from Arkansas,21
which indicates that tick bite accounts for the majority of
human exposures to F. tularensis in the state. The discrepant
result from the Illinois-Indiana-Kentucky-Tennessee area,
where county tularemia incidence was found to be negatively
associated with coverage by dry forest habitat suitable for
ticks, is intriguing. This may have resulted from a variety of
reasons including differences in species composition or abun-
County unit-based associations between percentage coverage by different habitat types and tularemia incidence class, 1990–2003, in the south-
central United States
(N ? 190 counties)
(N ? 275 counties)
(N ? 409 counties)
Suitable habitat for vector ticks
Dry forest—all types combined
Dry deciduous forest
Dry coniferous forest
Dry mixed forest
Partially suitable habitat for vector ticks
Grassland (including pasture)
Unsuitable habitat for vector ticks
Seasonally flooded habitats and wetlands
* Standardized classifications based on U.S. Geological Survey state Gap analysis projects.
† Association, based on ordinal logistic regression, classified as positive (P < 0.05), negative (P < 0.05), or none (P > 0.05).
‡ Based on 12 tularemia incidence classes: 0, 0.001–0.049, 0.050–0.099, 0.100–0.199, 0.200–0.299, 0.300–0.399, 0.400–0.499, 0.500–0.749, 0.750–0.999, 1.000–1.499, 1.500–1.999, and 2.000–6.500
cases per 100,000 person-years.
EISEN AND OTHERS
dance of vector ticks or vertebrate reservoirs or from trans-
mission to humans occurring more frequently through routes
other than tick bite (e.g., handling of lagomorphs). Alterna-
tively, forested habitats may be more important for tick sur-
vival, relative to grasslands, in the drier environments to the
west (Oklahoma-Kansas-Nebraska and the eastern part of
Arkansas-Missouri) than in moister areas to the east (Illinois-
Indiana-Kentucky-Tennessee). On the other hand, one can
speculate that intense enzootic transmission of F. tularensis
and elevated risk of human pathogen exposure is related to
presence of dry mixed deciduous/coniferous forest, which oc-
curs commonly in the Arkansas-Missouri tularemia focus but
not in the Illinois-Indiana-Kentucky-Tennessee area. Finally,
these differences could be related to the habitat associations
of F. tularensis Types A and B, such that Type A is more
likely to occur in dry forested habitat than Type B. Type A is
considered to be more strongly associated with ticks relative
to Type B,2,4thus supporting this hypothesis. Intriguingly,
78% of samples from tularemia patients originating from Ar-
kansas-Missouri and 94% of samples from Oklahoma-
Kansas-Nebraska were classified as Type A, whereas 76% of
samples from the Illinois-Indiana-Kentucky-Tennessee area
were classified as Type B.34These considerations highlight
the need for renewed studies comparing both the ecology and
Parameter estimates and likelihood ratio tests for variables included
in a multivariate logistic regression model for presence of areas
with tularemia risk in Arkansas-Missouri
Precipitation in July
Annual relative humidity
0.0124 0.14160.011 0.930
tularemia in Arkansas-Missouri (gray). Counties reporting cases dur-
ing 1990–2003 are displayed with hatched lines.
Counties with predicted risk of human exposure to
tularemia agent in Arkansas-Missouri (rose-colored). Counties re-
porting cases during 1990–2003 are shaded light blue. This figure
appears in color at www.ajtmh.org.
Areas with predicted risk of human exposure to the
epidemiology of tularemia across a habitat gradient ranging
from Oklahoma-Kansas to Arkansas-Missouri and Kentucky-
Although our predictive model for risk of exposure to the
tularemia agent in Arkansas-Missouri was based on relatively
coarse county-based incidence data, it clearly showed the po-
tential for GIS-based modeling of tularemia ecoepidemiology
and highlighted the need for further efforts based on more
fine-scale point, zip code, or census tract–based epidemiologic
data. Modern GIS techniques and an explosion in GIS-based
environmental data have opened new avenues for epidemio-
logic research. We now, however, are faced with the challenge
of improving the quality of the epidemiologic data available
for use in GIS-based modeling approaches.35Recent GIS-
based models for plague ecoepidemiology in the southwest-
ern United States36–38provide powerful examples of what can
be accomplished in terms of modeling of spatial risk of vector-
borne diseases when reliable data on pathogen exposure sites
are available. Furthermore, a recent study on Lyme disease in
California39showed the value of assessing risk based on com-
bined acarologic vector data and epidemiologic data and of
calculating and displaying disease incidence at zip code rather
than county scales. We expect similar future approaches for
tularemia to generate improved spatial risk models for this
disease and hope that our initial effort has adequately under-
scored the critical need for reliable fine-scale tularemia case
information including not only the address of residence of
afflicted persons but also evaluations of likely pathogen ex-
In the absence of exhaustive case studies, the address of
residence will likely remain the best estimate for a probable
point of pathogen exposure for tularemia cases in the United
States. Such data are, however, restricted to areas where
people reside. In contrast, risk models based on vector abun-
dance can be extrapolated to publicly owned lands where
people do not reside but where exposure could occur during
recreational activities.39–41Therefore, field-based studies on
vector ticks are also needed to improve the accuracy of spatial
risk assessments for exposure to F. tularensis. Vector abun-
dance data collected using systematic sampling methods could
be useful for generating acarologic risk models for key vector
species (A. americanum, D. variabilis) and elucidating envi-
ronmental correlates of tick abundance.40These models
could also be useful for assessing risk of other tick-borne
diseases occurring in the southcentral and southeastern
United States (e.g., human monocytic ehrlichiosis, Rocky
Mountain spotted fever, southern tick-associated rash ill-
ness).31,42–49Fine-scale models for spatial risk patterns will
help to target the use of limited tick control resources to areas
with especially high risk and inform the local medical com-
munity and public in such areas of the elevated risk of expo-
sure to F. tularensis.
Our predictive model of tularemia risk in Arkansas-
Missouri accurately classified counties reporting cases as con-
taining risk areas. User and producer accuracy for predicting
case occurrences were notably higher for the sub-county scale
model evaluation than for the county scale one. This suggests
that the model, which was developed based on associations
between 2 × 2-km scale GIS-based environmental data and
county scale tularemia incidence data, is informative at sub-
county scales. The accuracy of the sub-county scale predic-
tions will be examined in future studies by comparisons of
coverage of predicted risk areas and observed tularemia in-
cidence by census tract or zip code for Arkansas-Missouri.
Overall accuracy was reduced because a high proportion of
counties that did not report any tularemia cases during 1990–
2003 were identified by the model as containing risk habitat.
This type of error is consistent with infrequently occurring
diseases (Arkansas and Missouri combined for an average of
≈50 cases per year during the study period) and also was
observed in the above-mentioned plague risk model for the
southwestern United States.36–38Anomalies in tularemia oc-
currence with cases lacking from counties classified by the
model as solid risk and bordered on all sides by counties
reporting cases (Figure 5) could be related to either incom-
plete case reporting or persons with mild cases of tularemia
not seeking medical attention. Alternatively, the error could
result from behavioral factors associated with county-specific
risk of pathogen exposure (e.g., variation in the amount of
time spent outdoors in tick habitat or rabbit hunting) and not
considered in this model. Overall model accuracy was also
reduced when the model did not identify risk areas from a
county but cases were reported. This type of error may result
from travel-related exposure and underscores the importance
of the medical community routinely determining whether a
tularemia case likely was acquired in the peridomestic envi-
ronment, outside the peridomestic environment but within
the county of residence, or outside of the county of residence.
Misclassification of counties in northern Missouri could be
related to the distribution of F. tularensis Types A and B.
Recent molecular analysis of F. tularensis isolates from hu-
Sub-county scale validation results for multivariate logistic regression
model for presence of areas with tularemia risk based on build set
or validation set for Arkansas-Missouri
correct† Tularemia reported Not reported
No risk predicted
No risk predicted
* Probability cut-off value of P ? 0.30 was used to classify risk.
† User accuracy (commission error).
‡ Producer accuracy (omission error).
County-scale validation results for multivariate logistic regression
model for presence of counties with tularemia risk based on build
set or validation set for Arkansas-Missouri
correct† Tularemia reportedNot reported
No risk predicted
No risk predicted
* Probability cut-off value of P ? 0.7141 was used to classify risk.
† User accuracy (commission error).
‡ Producer accuracy (omission error).
EISEN AND OTHERS
mans showed that Type A is common in southern Missouri
and Arkansas, whereas Type B is most common in northern
Missouri.34It is possible that our model is predictive primarily
of the risk of exposure to tick-borne Type A and that cases
from northern Missouri were exposed to Type B.
Recent advances in genetic characterization of F. tularensis
have allowed for determination of not only Type A versus
Type B but also different subtypes of Type A that may differ
in their pathogenicity to humans.34,50–54This provides intrigu-
ing possibilities for expanding our initial studies on the
ecoepidemiology of tularemia to also include F. tularensis
types and subtypes with differing perceived transmission
routes to or pathogenicity in humans. For example, Type A is
considered to be associated with ticks and lagomorphs,
whereas Type B in addition to being detected from ticks also
is thought to be associated with water and rodents such as
beavers and voles.1,2,4,55,56Future studies are needed to de-
termine environmental correlates of Type A versus Type B
and, especially, subtypes of Type A with high pathogenicity in
Received August 17, 2007. Accepted for publication January 3, 2008.
Acknowledgments: The authors thank Karen Yates and Bao-Ping
Zhu of the Missouri Department of Health and Senior Services for
helpful comments on the manuscript.
Financial support: This study was funded by a grant from the Colo-
rado State University College of Veterinary Medicine and Biomedi-
cal Sciences to L. Eisen.
Authors’ addresses: Rebecca J. Eisen, Paul S. Mead, Andrew M.
Meyer, Liza E. Pfaff, Kristy K. Bradley, and Lars Eisen, Division of
Vector-Borne Infectious Diseases, National Center for Zoonotic,
Vector-Borne and Enteric Diseases, Centers for Disease Control and
Prevention, 3150 Rampart Road, Fort Collins, CO 80522, Tel: 970-
221-6408, Fax: 970-221-6476, E-mail: email@example.com.
Reprint requests: Rebecca J. Eisen, Division of Vector-Borne Infec-
tious Diseases, National Center for Zoonotic, Vector-Borne and En-
teric Diseases, Centers for Disease Control and Prevention, 3150
Rampart Road, Fort Collins, CO 80521, E-mail: firstname.lastname@example.org.
1. Hayes EB, 2005. Tularemia. Goodman JL, Dennis DT, Sonen-
shine DE, eds. Tick-Borne Diseases of Humans. Washington,
DC: ASM Press, 207–217.
2. Jellison WL, 1974. Tularemia in North America, 1930–1974. Mis-
soula, MT: University of Montana.
3. Ellis J, Oyston PCF, Green M, Titball RW, 2002. Tularemia. Clin
Microbiol Rev 15: 631–646.
4. Hopla CE, 1974. The ecology of tularemia. Adv Vet Sci Comp
Med 18: 25–53.
5. Choi E, 2002. Tularemia and Q fever. Med Clin North Am 85:
6. Centers for Disease Control, 2005. Summary of notifiable dis-
eases—United States, 2003. MMWR Morb Mortal Wkly Rep
7. Centers for Disease Control, 2002. Tularemia: United States,
1990–2000. MMWR Morb Mortal Wkly Rep 51: 181–184.
8. Centers for Disease Control, 2006. Summary of notifiable dis-
eases—United States, 2004. MMWR Morb Mortal Wkly Rep
9. Feldman KA, 2003. Tularemia. J Am Vet Med Assoc 222: 725–
10. Eisen L, 2007. A call for renewed research on tick-borne Fran-
cisella tularensis in the Arkansas-Missouri primary national fo-
cus of tularemia in humans. J Med Entomol 44: 389–397.
11. Goddard J, McHugh CP, 1990. Impact of a severe tick infestation
at Little Rock AFB, Arkansas, on Volant Scorpion military
training. Mil Med 155: 277–280.
12. Hair JA, Howell DE, 1970. Lone Star Ticks. Their Biology and
Control in Ozark Recreation Areas. Oklahoma Agric. Exp. Stn.
13. Assal N, Blenden DC, Price ER, 1967. Epidemiologic study of
human tularemia reported in Missouri, 1949–65. Public Health
Rep 82: 627–632.
14. Assal NR, Lindeman RD, Carpenter RL, 1968. Epidemiologic
study on reported human tularemia in Oklahoma, 1944–65. J
Okla State Med Assoc 61: 120–124.
15. Bost RB, Percefull SC, Leming HE, 1948. Tularemia in the
Ozarks region. JAMA 137: 352–354.
16. Callaway GD, Peterson SS, Good JT, 1954. Tularemia in South-
west Missouri; a report and discussion of seventy-eight cases.
Mo Med 51: 906–909.
17. Corwin WC, Stubbs SP, 1952. Further studies on tularemia in the
Ozarks. Review of forty-four cases during a three-year period.
JAMA 149: 343–345.
18. Cox SK, Everett ED, 1981. Tularemia, an analysis of 25 cases. Mo
Med 78: 70–74.
19. Evans ME, Gregory DW, Schaffner W, McGee ZA, 1985. Tula-
remia: a 30-year experience with 88 cases. Medicine 64: 251–
20. Lopez CE, Kornblatt AN, Sikes RK, Hanes OE, 1982. Tularemia:
review of eight cases of tick-borne infection and the epidemi-
ology of the disease in Georgia. South Med J 75: 404–407.
21. McChesney TC, Narain J, 1983. A five-year evaluation of tula-
remia in Arkansas. J Ark Med Soc 80: 257–262.
22. Pullen RL, Stuart BM, 1945. Tularemia. Analysis of 225 cases.
JAMA 129: 495–500.
23. Taylor JP, Istre GR, McChesney TC, Satalowich FT, Parker RL,
McFarland LM, 1991. Epidemiologic characteristics of human
tularemia in the southwest-central states, 1981–1987. Am J
Epidemiol 133: 1032–1038.
24. Washburn AM, Tuohy JM, 1949. The changing picture of tula-
remia transmission in Arkansas. A study of 704 case histories.
South Med J 42: 60–62.
25. Rohrbach BW, Westerman E, Istre GR, 1991. Epidemiology and
clinical characteristics of tularemia in Oklahoma, 1979 to 1985.
South Med J 84: 1091–1096.
26. Centers for Disease Control. 1997. Case definitions for infectious
conditions under public health surveillance. Centers for Dis-
ease Control and Prevention. MMWR Recomm Rep 46: 1–55.
27. Centers for Disease Control. 2002. Tularemia—United States,
1990–2000. MMWR Morb Mortal Wkly Rep 51: 182–184.
28. Sonenshine DE, Atwood EL, Lamb JT Jr, 1966. The ecology of
ticks transmitting Rocky Mountain spotted fever in a study
area in Virginia. Ann Entomol Soc Am 59: 1234–1262.
29. Sonenshine DE, Levy GF, 1972. Ecology of the American dog
tick, Dermacentor variabilis, in a study area in Virginia. 2.
Distribution in relation to vegetative types. Ann Entomol Soc
Am 65: 1175–1182.
30. Sonenshine DE, 1993. Biology of Ticks. Volume 2. New York:
Oxford University Press.
31. Yabsley MJ, Wimberly MC, Stallknecht DE, Little SE, Davidson
WR, 2005. Spatial analysis of the distribution of Ehrlichia chaf-
feensis, causative agent of human monocytotropic ehrlichiosis,
across a multi-state region. Am J Trop Med Hyg 72: 840–850.
32. Sall J, Creighton L, Lehman A, 2005. JMP Start Statistics. Third
Edition. Belmont, CA: Brooks/Cole.
33. Hopla CE, 1962. Ixodes scapularis as a vector of tularemia or-
ganisms in the southern United States. Verh. 11 Int. Kongr.
Ent. Vienna, August 17–25, 1960.
34. Staples JE, Kubota KA, Chalcraft LG, Mead PS, Petersen JM,
2006. Epidemiologic and molecular analysis of human tulare-
mia, United States, 1964–2004. Emerg Infect Dis 12: 1113–1118.
35. Eisen L, Eisen RJ, 2007. Need for improved methods to collect
and present spatial epidemiologic data for vectorborne dis-
eases. Emerg Infect Dis 13: 1816–1820.
36. Eisen RJ, Enscore RE, Biggerstaff BJ, Reynolds PJ, Ettestad P,
Brown T, Pape J, Tanda D, Levy CE, Engelthaler DM, Cheek
J, Bueno R Jr, Targhetta J, Montenieri JA, Gage KL, 2007.
Human plague in the southwestern United States, 1957–2004:
spatial models of elevated risk of human exposure to Yersinia
pestis. J Med Entomol 44: 530–537.
37. Eisen RJ, Glass GE, Eisen L, Cheek J, Enscore RE, Ettestad P,
Gage KL, 2007. A spatial model of shared risk for plague and
hantavirus pulmonary syndrome in the southwestern United
States. Am J Trop Med Hyg 77: 999–1004.
38. Eisen RJ, Reynolds PJ, Ettestad P, Brown T, Enscore RE, Big-
gerstaff BJ, Cheek J, Bueno R, Targhetta J, Montenieri JA,
Gage KL, 2007. Residence-linked human plague in New
Mexico: a habitat-suitability model. Am J Trop Med Hyg 77:
39. Eisen RJ, Lane RS, Fritz CL, Eisen L, 2006. Spatial patterns of
Lyme disease risk in California based on disease incidence
data and modeling of vector-tick exposure. Am J Trop Med
Hyg 75: 669–676.
40. Eisen RJ, Eisen L, Lane RS, 2006. Predicting density of Ixodes
pacificus nymphs in dense woodlands in Mendocino County,
California, based on geographic information systems and re-
mote sensing versus field-derived data. Am J Trop Med Hyg
41. Eisen L, Meyer AM, Eisen RJ, 2007. Climate-based model pre-
dicting acarological risk of encountering the human-biting
adult life stage of Dermacentor andersoni (Acari: Ixodidae) in
a key habitat type in Colorado. J Med Entomol 44: 694–704.
42. Barbour AG, Maupin GO, Teltow GJ, Carter CJ, Piesman J,
1996. Identification of an uncultivable Borrelia species in the
hard tick Amblyomma americanum: possible agent of a Lyme
disease-like illness. J Infect Dis 173: 403–409.
43. Chapman AS, Murphy SM, Demma LJ, Holman RC, Curns AT,
McQuiston JH, Krebs JW, Swerdlow DL, 2006. Rocky Moun-
tain spotted fever in the United States, 1997–2002. Vector
Borne Zoonotic Dis 6: 170–178.
44. Dalton MJ, Clarke MJ, Holman RC, Krebs JW, Fishbein DB,
Olson JG, Childs JE, 1995. National surveillance for Rocky
Mountain spotted fever, 1981–1992: epidemiologic summary
and evaluation of risk factors for fatal outcome. Am J Trop
Med Hyg 52: 405–413.
45. Demma LJ, Holman RC, McQuiston JH, Krebs JW, Swerdlow
DL, 2005. Epidemiology of human ehrlichiosis and anaplas-
mosis in the United States, 2001–2002. Am J Trop Med Hyg 73:
46. James AM, Liveris D, Wormser GP, Schwartz I, Montecalvo
MA, Johnson BJB, 2001. Borrelia lonestari infection after a
bite by an Amblyomma americanum tick. J Infect Dis 183:
47. McQuiston JH, Paddock CD, Holman RC, Childs JE, 1999. The
human ehrlichioses in the United States. Emerg Infect Dis 5:
48. Paddock CD, Childs JE, 2003. Ehrlichia chaffeensis: a prototypi-
cal emerging pathogen. Clin Microbiol Rev 16: 37–64.
49. Varela AS, Luttrell MP, Howerth EW, Moore VA, Davidson
WR, Stallknecht DE, Little SE, 2004. First culture isolation of
Borrelia lonestati, putative agent of southern tick-associated
rash illness. J Clin Microbiol 42: 1163–1169.
50. Farlow J, Wagner DM, Dukerich M, Stanley M, Chu M, Kubota
K, Petersen J, Keim P, 2005. Francisella tularensis in the
United States. Emerg Infect Dis 11: 1835–1841.
51. Kugeler KJ, Gurfield N, Creek JG, Mahoney KS, Versage JL,
Petersen JM, 2005. Discrimination between Francisella tula-
rensis and Francisella-like endosymbionts when screening ticks
by PCR. Appl Environ Microbiol 71: 7594–7597.
52. Johansson A, Farlow J, Larsson P, Dukerich M, Chambers E,
Byström M, Fox J, Chu M, Forsman M, Sjöstedt A, Keim P,
2004. Worldwide genetic relationships among Francisella tula-
rensis isolates determined by multiple-locus variable-number
tandem repeat analysis. J Bacteriol 186: 5808–5818.
53. Farlow J, Smith KL, Wong J, Abrams M, Lytle M, Keim P, 2001.
Francisella tularensis strain typing using multiple-locus, vari-
able-number tandem repeat analysis. J Clin Microbiol 39:
54. Goethert HK, Shani I, Telford SR III, 2004. Genotypic diversity
of Francisella tularensis infecting Dermacentor variabilis ticks
on Martha’s Vineyard, Massachusetts. J Clin Microbiol 42:
55. Markowitz LE, Hynes NA, de la Cruz P, Campos E, Barbaree
JM, Plikaytis BD, Mosier D, Kaufmann AF, 1985. Tick-borne
tularemia. An outbreak of lymphadenopathy in children.
JAMA 254: 2922–2925.
56. Schmid GP, Kornblatt AN, Connors CA, Patton C, Carney J,
Hobbs J, Kaufmann AF, 1983. Clinically mild tularemia asso-
ciated with tick-borne Francisella tularensis. J Infect Dis 148:
EISEN AND OTHERS