American Journal of Epidemiology
Copyright ª 2005 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved; printed in U.S.A.
Vol. 163, No. 2
DOI: 10.1093/aje/kwj023
Advance Access publication November 23, 2005
Practice of Epidemiology
Geographic Prediction of Human Onset of West Nile Virus Using Dead Crow
Clusters: An Evaluation of Year 2002 Data in New York State
Glen D. Johnson1,2, Millicent Eidson1,3, Kathryn Schmit1, April Ellis1, and Martin Kulldorff4
1
Zoonoses Program, New York State Department of Health, Albany, NY.
Department of Environmental Health Sciences, School of Public Health, University at Albany, Albany, NY.
3
Department of Epidemiology, School of Public Health, University at Albany, Albany, NY.
4
Department of Ambulatory Care and Prevention, Harvard Medical School and Harvard Pilgrim Health Care, Boston, MA.
2
Received for publication December 14, 2004; accepted for publication September 2, 2005.
The risk of becoming a West Nile virus case in New York State, excluding New York City, was evaluated for persons whose town of residence was proximal to spatial clusters of dead American crows (Corvus brachyrhynchos).
Weekly clusters were delineated for June–October 2002 by using both the binomial spatial scan statistic and
kernel density smoothing. The relative risk of a human case was estimated for different spatial-temporal exposure
definitions after adjusting for population density and age distribution using Poisson regression, adjusting for week
and geographic region, and conducting Cox proportional hazards modeling, where the week that a human case
was identified was treated as the failure time and baseline hazard was stratified by region. The risk of becoming
a West Nile virus case was positively associated with living in towns proximal to dead crow clusters. The highest
risk was consistently for towns associated with a cluster in the current or prior 1–2 weeks. Weaker, but positive
associations were found for towns associated with a cluster in just the 1–2 prior weeks, indicating an ability to
predict onset in a timely fashion.
arboviruses; geographic information systems; Poisson distribution; population surveillance; proportional hazards
models; space-time clustering; West Nile virus; zoonoses
Abbreviation: WNV, West Nile virus.
Since it was first detected in New York City in 1999, West
Nile virus (WNV) has become an expanding pandemic in
the Western hemisphere (1), thus requiring active surveillance to identify geographic areas of high risk for human
infection. This flavivirus is sustained in various avian species and is transmitted by mosquitoes among birds and from
birds to mammals when the cycle amplifies (2). With WNV,
human infection appears to be spatially-temporally associated with unusual bird die-offs, particularly among American crows (Corvus brachyrhynchos) (3–5).
New York State has responded by exploiting its Health
Information Network, a secure, Web-based, enterprise-wide
information infrastructure (6), to maintain real-time moni-
toring data related to WNV. Data include dead bird locations
reported by the public, mandatory reporting of human cases,
and laboratory tests for mosquitoes, birds, horses, and other
animals.
Although laboratory results are necessary to confirm the
presence of WNV, the average time of 23 days (based on the
2002 WNV season) for collection, shipment, processing,
and testing of dead bird specimens reduces the real-time
utility of these results for guiding prevention and control
activities prior to human onset. In contrast, monitoring of
crow deaths requires only a phone call and some data entry.
Crows exhibit a high case-fatality rate, dying soon after
infection, and they are both easily identified and fairly
Correspondence to Dr. Millicent Eidson, Zoonoses Program, New York State Department of Health, 621 Corning Tower, Empire State Plaza,
Albany, NY 12237 (e-mail: [email protected]).
171
Am J Epidemiol 2006;163:171–180
172 Johnson et al.
commonly distributed throughout the state (7). Simple maps
of dead crow locations have proven valuable for identifying
areas of high WNV activity (8), regardless of observation
bias in such maps. Indeed, retrospective studies have revealed a positive association between county-level dead
crow density (number per square mile) and the number of
human cases for the years 2000 (9) and 2001 (10).
The initial studies cited above indicated that dead crow
density may provide a critical real-time sentinel for WNV
activity; however, geographic resolution finer than counties
is required for community-level prevention and control activities (10). Furthermore, the surveillance system would
benefit from a more objective method for identifying clusters of dead birds, or dead crows in particular.
To this end, two different methods were applied in New
York City in the year 2001 to identify bird clusters that are
statistically significant with respect to spatially random expectation (11, 12). Both approaches analyzed the locations
of dead birds, excluding pigeons and unknown species,
which were geocoded to census tracts. One approach (11)
applied the spatial scan statistic (13) daily, using the Bernoulli model that compared dead bird counts after confirmed
WNV activity (cases) with preoutbreak dead bird counts
(controls) in each tract. The other approach (12), based on
work by Knox (14), applied a contingency table analysis to
detect significant space-time interaction of dead bird counts
based on predefined definitions of ‘‘closeness’’ in both space
and time. After calibration with previous data for defining
‘‘closeness,’’ the test was applied to each grid cell of a surface tessellation to estimate the probability of nonrandom
space-time interaction of bird deaths. The result is a continuous surface of probability estimates, which helps focus
areas of amplified WNV activity. Both the daily spatial scan
statistic and the space-time Knox test identified significant
dead bird clusters in similar locations that also captured the
residences of five of seven diagnosed human cases in 2001.
Aside from New York City, the rest of New York State is
analyzed for WNV activity each season by using data from
the Health Information Network (6). Both the spatial scan
statistic (13) and kernel density smoothing (15) are applied
weekly to reported dead crow locations. Resulting maps are
provided through the Health Information Network to help
county health departments prioritize prevention and control
efforts. These maps visually appear to effectively indicate
WNV activity and even predict human onset; however, just
like all studies discussed above, a formal statistical evaluation of the association between dead bird/crow clusters and
human case onset has been lacking.
A recent retrospective study of WNV in Chicago, Illinois,
for the year 2002 did show significant incidence ratios of 3.0
for WNV infection and 2.3 for WNV meningo-encephalitis
for residents in high crow-mortality areas relative to those
outside these areas (16). Areas whose kernel density estimates were in the top 90 percent (density >0.10) were chosen as ‘‘high,’’ and the whole season’s data were pooled.
Age was the only potential confounder evaluated and apparently did not yield a significant or confounding effect.
In this paper, we quantify the association between dead
crow clusters and human onset of WNV in New York State,
outside of New York City, for the year 2002, while also
considering the potential confounders or effect modifiers
of human population density and age distribution. The effects of time are evaluated two ways: as an implicit covariate
in a Poisson regression and as the ‘‘waiting time’’ until onset
in a Cox proportional hazards model.
MATERIALS AND METHODS
Data
Dead crow locations. Weekly reports of dead crow sightings were downloaded from the dead bird surveillance database on New York State’s Health Information Network (6)
for June 2–October 12, 2002, corresponding to weeks 23–
41, as recognized by the US Centers for Disease Control and
Prevention. Dead bird locations are reported by the public
through either their county health departments or a dead bird
hotline maintained by the New York office of the US Department of Agriculture Wildlife Service. Counties outside
of New York City were assigned to six regions, as indicated
in figure 1, for managing geocoding activity and performing
region-specific cluster analysis. Most dead bird locations
were geocoded to an actual street address by using MapMarker (17) supplemented by manual geocoding based on
investigative information for those reports that could not
automatically be geocoded to a street address. Only those
dead crow reports that were available on the Health Information Network in real time and were used for mapping in
the subsequent week were utilized for this analysis. Dead
crow reports entered into the Health Information Network at
later dates were not available for mapping and were not
used.
Human cases. Human cases were those who met the
Centers for Disease Control and Prevention’s case definition
of laboratory-confirmed or probable WNV-associated fever
or neurologic disease (18). Of 53 cases reported in New
York State outside of New York City for 2002, we analyzed
data for 50 cases whose reported onset date was within 2
weeks after dead crow mapping ended for the season.
Human population characteristics. Data from the year
2000 US Census were used for population counts by age
in county subdivisions, which in New York State are called
townships. Our study included 1,006 subdivisions outside of
New York City.
Cluster detection
Spatial cluster analysis was performed weekly in each of
the six regions shown in figure 1 for the dead crow reports of
the particular week. Two methods were used for identifying
clusters—the spatial scan statistic (13) and kernel density
smoothing (15). An example of results for these two methods is shown for the Long Island region in figure 2, and
a detailed discussion follows.
Spatial scan statistic. Using SaTScan 3.0 software (19),
we applied the binomial model of the purely spatial scan
statistic (11) to each of the six regions separately. With dead
crow locations assigned to census tracts, circular search
windows started with individual tracts and expanded to include neighboring tracts until a maximum of 15 percent of
Am J Epidemiol 2006;163:171–180
Prediction of West Nile Virus Using Crow Clusters
173
FIGURE 1. The six regions of New York State used for separate spatial analyses of dead crow clusters during the week of September 1–7, 2002.
all crows, including the current week’s cases and baseline
controls, within the region was reached. This ‘‘15 percent’’
criterion is based on previous research (20). Given the dead
birds reported within a monitoring week (cases) and during
a control period prior to the WNV season (controls), the
scan statistic calculates a likelihood ratio for each search
window, which is proportional to
c c n cnc C c Cc
n
n
N n
ðN nÞ ðC cÞ ðNnÞðCcÞ
3
N n
for c cases and n total cases and controls inside the search
window, and C cases and N total cases and controls throughout the region, including within the search window. Each
observed likelihood ratio was then compared with 999
Monte Carlo simulations from the null model to evaluate
significance. Two separate sets of spatial scan statistic clusters were then retained: those significant at the p < 0.05
level and those significant at the p < 0.01 level. The control
period was from January 1 until 3 weeks prior to the week of
analysis, allowing at least a 2-week buffer between crow
deaths in the absence of viral activity (controls) and those
in its presence (cases). When the first laboratory-positive
bird was reported in a region, the control period ended with
that bird’s found date for all subsequent analyses.
Kernel density smoothing. Using Vertical Mapper (21)
within MapInfo’s geographic information system software
(22), we estimated dead crow density for cells of a raster
grid based on the sum of kernel functions within each
Am J Epidemiol 2006;163:171–180
cell, standardized to the real unit interval from 0 to 1. The
Epanechnikov kernel estimate was chosen to weight neighbors as a parabolically decreasing function of distance until
the edge of the search radius was reached. An adaptive
bandwidth was allowed to change for each grid cell to include a specified number of closest points, which were in
turn chosen each week according to an adaptive kernel
smoothing key we created. This key related the overall number of dead crow locations to an associated number of closest points. Areas whose cells had densities above 0.5 were
transformed by a contouring procedure into polygons to
delineate clusters. Experience shows that clusters become
too diffuse and noninformative as densities drop below 0.5.
Exposure definition
Exposure was defined as being close in space and time to
a dead crow cluster. Spatial association with dead crow
clusters was determined by using a geographic information
system (22) according to the following protocol. Census
tracts within clusters were selected, and a 1-mile buffer
(1 mile ¼ 1.609 km) was created around the common outer
boundary of these tracts, since the principal mosquito carrier
in the eastern United States, Culex pipiens, is known to fly
up to 1 mile (23). Each of the 1,006 towns outside of New
York City was then identified as being ‘‘in or adjacent to’’
a cluster in a particular week if the town boundary intersected a cluster or a buffer area for that week.
Temporal associations were designed to address particular questions. To evaluate whether there is a spatial relation
between dead crows and human cases, we considered towns
174 Johnson et al.
FIGURE 2. Dead crow locations and cluster analyses for Long Island, New York, during the week of September 1–7, 2002: kernel density
smoothing (A) and spatial scan statistic clusters (B) (note that all clusters significant at the p < 0.05 level were also significant at p < 0.01).
exposed during the ‘‘current plus any prior week’’ of the
season, and we conducted the more focused evaluation of
exposure during the ‘‘current plus prior 1–2 weeks’’ based
on the incubation period of approximately 3–14 days in
humans (24). To evaluate the ability of recent dead crow
clusters to predict human onset, we considered towns exposed during only the prior 1–2 weeks.
Estimating relative risk
The period used for estimating risk was the week of July
28, when the first human case was recorded, to the week of
October 5, when the last human case was recorded for which
crow surveillance data were also available within the pre-
vious 2 weeks. Risk estimates therefore represent the 2002
WNV season in New York, conditional on occurrence of the
first human case. Poisson regression and Cox proportional
hazards modeling were applied separately to estimate the
relative risk of a human case during the 10-week season
using different, but related methods.
Using Poisson regression (25), we regressed the number
of cases in each of n ¼ 10,060 unique combinations of the
1,006 towns and 10 weeks on exposure after offsetting the
total population and adjusting for possible confounders and
other covariates that may explain geographic variability of
human onset. Exposure was defined by a binary indicator
variable (exposed/unexposed). Specific covariates were
1) region, as indicated in figure 1; 2) Centers for Disease
Am J Epidemiol 2006;163:171–180
Prediction of West Nile Virus Using Crow Clusters
Control and Prevention week (category valued); 3) different
transformations of town population density (people per
square mile); and 4) proportion of the town population over
age 50 years, since persons older than age 50 years are at
higher risk of symptomatic meningo-encephalitis (26). Both
the population density and the age distribution were evaluated as potential confounders or effect modifiers with respect to exposure. Poisson modeling was performed with
SAS 8.02 software (SAS Institute, Inc., Cary, North Carolina) by using the GENMOD procedure.
Realizing that the time until human onset of WNV is
essentially a ‘‘waiting time’’ situation, we turned to the
theory of Cox proportional hazards modeling (27). The instantaneous disease rate for individual i at any observed time
t, defined as the hazard function hi(t), is modeled as hi(t) ¼
exp(b#x)h0(t), where x and b are vectors of explanatory
covariates and their coefficients, respectively, and h0(t) is
the hazard corresponding to ‘‘baseline’’ values of the covariates x. For a binary exposure variable (xE ¼ 1 if exposed
^E), proand 0 otherwise), the estimated hazard ratio, exp(b
vides an estimate of relative risk after adjusting for the other
covariates.
Exposure was modeled as a time-dependent variable that
could arise and go away repeatedly throughout the 10-week
study period. The effect of geographic region, as depicted in
figure 1, was evaluated separately as both an implicit covariate and a stratifier, where the baseline hazard, h0(t), was
allowed to be unique for each region. The other covariates
mentioned in the Poisson regression discussion were also
examined to help identify potential confounders and effect
modifiers when estimating hazard ratios.
Each of the 1,006 towns in this study was weighted by its
total population, representing the approximately 11 million
people not reported as being infected with WNV. They were
treated as censored observations because they were not infected/diagnosed within the observation period. Each of the
50 cases was treated as a town with a population of size one.
Town-level demographic information was then associated
with cases and noncases according to their town of residence.
Proportional hazards modeling was performed with the
PHREG procedure in SAS 8.02 software, with population
weights specified by the FREQ statement, and Efron’s approximation to the partial likelihood (28) was chosen for
handling tied onset times. Although observations were made
in discrete time (weekly), the process of human infection
and disease development occurs in real time. For the methods available for tied ‘‘failure’’ times of a real-time process,
Efron’s method was chosen because it provides the closest
approximation to the actual likelihood (29).
RESULTS
The distributions of dead crows and human WNV cases
are presented in table 1 by region and week. This table
shows the rise and fall of WNV activity statewide, where
the dead crow reports peak 1 week prior to the peak in human cases. The number of human cases and the population
size exposed to dead crow clusters are reported in table 2 for
the different exposure definitions. Data from tables 1 and 2
Am J Epidemiol 2006;163:171–180
175
can be used to calculate crude relative risks. Estimates of the
relative risk for human WNV cases associated with the various spatial-temporal exposure definitions are presented in
tables 3 and 4 for Poisson and proportional hazards regression, respectively.
For Poisson regression, results are presented for the unadjusted model to show crude relative risks, along with the
model that adjusts for region and week only and the full
model including adjustments for human demographics.
Note that interaction terms were never significant, indicating a lack of effect modification by any covariates. For the
full models reported in table 3 that evaluate spatial scan
statistic clusters, each covariate was significant ( p < 0.01),
after we adjusted for all other covariates, for all three temporal definitions of exposure. When kernel density clusters
were evaluated, results were similarly significant for exposure 1–2 weeks prior. However, for the other two temporal
definitions, significance of the human demographic covariates was reduced (0.01 < p < 0.05), and population density
squared was rendered insignificant.
For proportional hazards modeling, week is no longer
a covariate because it becomes the response variable. Region was used as a stratifier, thus allowing the baseline
hazard to vary among regions. The importance of stratifying
by region was illustrated by unstratified models in which
region was treated as an implicit covariate, resulting in
much lower and less significant relative risks of exposure.
The remaining covariates for population density and age
distribution had the same effects as with Poisson regression;
therefore, table 4 includes the full model with all covariates,
along with the crude model (no stratification or covariate
adjustment) and the crude-stratified model (no covariate
adjustment).
DISCUSSION
Our findings indicate that the risk of human WNVassociated fever or neurologic disease is higher for persons
living in towns in or adjacent to clusters of dead crows than
for persons not living adjacent to such clusters. Most importantly, there is an increased risk with exposure during the 1–2
weeks prior, providing evidence that dead crow clusters can
predict human onset in a timely fashion. The predictive ability was observed even after we adjusted for human population density, and clusters delineated by the scan statistic
according to criteria of p < 0.05 and p < 0.01 had a positive
association that was significant at the levels of p ¼ 0.03 and
p ¼ 0.02, respectively, when evaluated by Poisson regression. This finding indicates that clusters of dead crow reports
may be used to predict human WNV in areas outside of New
York State, where mosquito-infested and WNV-infected
areas may be more rural.
Of the two methods used for delineating spatial crow
clusters, the scan statistic has the distinct advantage of providing statistical significance, thus allowing a more objective basis for identifying clusters. Furthermore, clusters
delineated by the scan statistic are much less sensitive to
confounding by human population density when compared
with kernel density clusters, as shown in tables 3 and 4.
176 Johnson et al.
TABLE 1. Distribution of dead crow reports (no.) and of human cases of West Nile virus (no.), by week in 2002 and region, New York State (excluding New York City)*
Control period
crow data
Week of the West Nile virus seasony
Total
Period
end date
No. of
crows
7/14–7/20
7/21–7/27
7/28–8/3
8/4–8/10
8/11–8/17
8/18–8/24
8/25–8/31
9/1–9/7
9/8–9/14
9/15–9/21
9/22–9/28
9/29–10/5
Crow reports
North
July 9
167
14
31
16
33
53
26
47
67
43
39
24
19
412
West
June 15
202
61
81
73
130
349
518
582
483
461
252
204
155
3,349
Central
June 21
168
23
33
32
71
115
221
447
535
329
275
176
103
2,360
Capital
May 1
119
10
10
24
52
52
73
89
72
48
28
28
24
510
South
April 25
118
84
77
86
123
160
282
238
227
166
103
103
55
1,704
Long Island
May 27
15
61
53
102
71
126
136
172
166
88
58
34
8
1,075
789
253
285
333
480
855
1,256
1,575
1,550
1,135
755
569
364
9,410
North
0
0
0
0
0
1
0
0
0
0
1
West
0
0
0
1
2
4
0
3
1
1
12
Central
0
0
1
0
3
3
2
2
2
0
13
Capital
0
1
0
0
0
0
1
0
1
0
3
Statewide
Human cases
Am J Epidemiol 2006;163:171–180
South
0
0
0
1
0
0
0
0
2
0
3
Long Island
1
1
3
2
2
5
4
0
0
0
18
Statewide
1
2
4
4
7
13
7
5
6
1
50
* The number of dead crows serving as controls for the binomial spatial scan statistic is shown, along with the final end date of the control period. For the West Nile virus season, dead crow
reports are for each week starting with 2 weeks prior to the first human case, and the 50 human cases are reported for the week of disease onset.
y Weeks are formatted as follows, for example: 7/14–7/20, July 14 to July 20.
Prediction of West Nile Virus Using Crow Clusters
177
TABLE 2. Number of human cases of West Nile virus and population size (in millions) exposed to clusters of dead crows, according
to different exposure definitions, during Centers for Disease Control and Prevention weeks 31–40 (July 28–October 5) in 2002,
New York State (excluding New York City)*
Weeky
7/28–8/3
8/4–8/10
8/11–8/17
8/18–8/24
8/25–8/31
9/1–9/7
9/8–9/14
9/15–9/21
9/22–9/28
9/29–10/5
SaTScanz p < 0.05
1 or 2 weeks prior
Cases
0
2
4
3
3
5
4
1
5
0
Population
2.7
5.1
5.6
3.8
2.3
4.0
4.8
4.9
3.6
4.1
Current or 1–2 weeks prior
Cases
1
2
4
3
4
5
6
1
5
0
Population
6.0
5.7
5.6
4.6
4.0
5.1
5.1
5.2
4.2
4.8
Any prior week
Cases
1
2
4
4
5
Population
6.4
7.2
7.6
7.6
7.8
10
8.1
6
2
5
0
8.1
8.1
8.1
8.1
SaTScan p < 0.01
1 or 2 weeks prior
Cases
0
2
3
3
1
5
3
0
5
0
Population
1.1
4.1
4.2
3.1
1.3
3.0
3.4
3.3
3.1
3.0
Current or 1–2 weeks prior
Cases
1
2
3
3
4
5
4
1
5
0
Population
4.7
4.3
4.3
3.5
3.0
4.1
3.6
4.2
3.4
3.4
Any prior week
Cases
1
2
4
4
2
9
6
1
5
0
Population
4.2
6.1
6.1
6.1
6.1
6.4
6.4
6.4
6.4
6.7
Kernel density
1 or 2 weeks prior
Cases
0
2
2
3
2
6
5
0
3
0
Population
1.9
2.6
2.7
2.3
2.8
4.0
4.4
3.2
4.1
4.4
Current or 1–2 weeks prior
Cases
0
2
2
4
6
9
5
2
4
0
Population
2.8
2.8
2.7
3.2
4.4
4.5
4.4
4.6
4.4
3.3
Any prior week
Cases
1
2
2
4
6
Population
5.9
6.0
6.0
6.0
6.0
11
6.2
7
3
6
0
6.3
6.3
6.7
6.8
* The number of unexposed cases each week can be obtained by subtracting from the total cases provided in table 1. The unexposed
population can be obtained by subtracting from the total population outside of New York City, which equals 10,968,179.
y Weeks are formatted as follows, for example: 7/28–8/3, July 28 to August 3.
z Refer to reference 19 for more information about this software.
A trade-off of such a parametric test is that it requires distributional assumptions, although these assumptions are reasonable. A distinct disadvantage of the binomial model of
the scan statistic, as applied in this paper, is that it depends
on a control population of dead crows. The number of controls is determined by factors that vary both within and
between regions, such as surveillance effort, public interest
in reporting, and the location of susceptible birds. For example, Long Island had a substantially smaller control
group than the other regions (table 1); therefore, clusters
may depend more on the location of control birds than on
the location of case birds in this region. However, it is
Am J Epidemiol 2006;163:171–180
worthwhile to note that, of the 18 cases on Long Island,
the scan statistic did predict (1–2-week prior analyses) 10
of them, compared with nine predicted by kernel density.
Another potential disadvantage of the scan statistic in general is that it evaluates artificial circular clusters. Therefore,
this method could prove to be more valuable if the potential
cluster boundaries were allowed to be more flexible, driven
by the data, reflecting more natural cluster shapes (30–32).
Kernel density delineation of clusters has the advantage
of being a rather simple, nonparametric smoothing method
that does not depend on a control population. However, as
noted above, it is very sensitive to confounding by human
178 Johnson et al.
TABLE 3. Risk, estimated by Poisson regression modeling, of human cases of West Nile virus in 2002 in
New York State (excluding New York City) for various definitions of exposure to dead crow clusters relative
to being unexposed
Clustering method
Spatial scan statistic
Temporal criteria and adjustments
Kernel density
p < 0.05
RR*
p < 0.01
95% CI*
RR
95% CI
RR
95% CI
1–2 weeks prior
None
1.97
1.13, 3.44
2.29
1.31, 4.00
2.59
1.49, 4.50
Region, week
1.99
1.09, 3.62
2.14
1.13, 4.04
2.22
1.22, 4.06
Fully
1.94
1.06, 3.53
2.11
1.15, 3.87
1.25
0.65, 2.42
Current week or 1–2 weeks prior
None
2.50
1.38, 4.52
3.02
1.71, 5.35
3.95
2.18, 7.16
Region, week
2.63
1.37, 5.05
3.32
1.70, 6.50
3.61
1.91, 6.81
Fully
2.41
1.26, 4.63
2.91
1.52, 5.57
2.15
1.03, 4.48
Any prior week
None
1.50
0.77, 2.94
1.70
0.94, 3.07
4.00
1.88, 8.52
Region, week
1.13
0.55, 2.31
1.40
0.68, 2.87
3.85
1.75, 8.48
Fully
1.41
0.67, 2.94
1.37
0.68, 2.76
2.23
0.91, 5.49
* RR, relative risk; CI, confidence interval.
y Adjusted for region, week, town population density, density squared, and proportion of town population aged
>50 years.
population density since it uses the size of the geographic
area, rather than a set of control birds, in the denominator.
Of course, a major disadvantage is that choosing a critical
density for delineating clusters is arbitrary. Kernel density
smoothing will always help to visualize spatial point densities, but it should be supplemented by quantitative methods
such as the scan statistic when objective cluster delineation
is necessary.
TABLE 4. Risk, estimated as a hazard ratio by Cox proportional hazards modeling, of human cases of
West Nile virus in 2002 in New York State (excluding New York City) for various definitions of exposure to
dead crow clusters relative to being unexposed
Clustering method
Spatial scan statistic
Kernel density
Temporal criteria and adjustments
p < 0.05
RR*
95% CI*
p < 0.01
RR
95% CI
RR
95% CI
1–2 weeks prior
None
1.70
0.93, 3.11
1.91
1.03, 3.53
2.49
1.37, 4.53
Stratified
1.65
0.82, 3.33
1.40
0.65, 3.03
2.51
1.28, 4.92
Stratified-adjustedy
1.47
0.74, 2.93
1.35
0.64, 2.84
1.61
0.79, 3.29
Current week or 1–2 weeks prior
None
2.33
1.24, 4.35
2.70
1.48, 4.94
3.70
1.95, 7.02
Stratified
2.18
1.07, 4.44
2.25
1.04, 4.89
3.65
1.81, 7.33
Stratified-adjustedy
1.87
0.93, 3.76
1.97
0.93, 4.19
2.34
1.09, 5.02
Any prior week
None
1.27
0.63, 2.57
2.14
1.15, 4.00
3.38
1.57, 7.27
Stratified
0.95
0.44, 2.03
1.87
0.88, 3.97
2.95
1.31, 6.64
Stratified-adjustedy
1.11
0.51, 2.42
1.61
0.77, 3.39
1.61
0.67, 3.90
* RR, relative risk; CI, confidence interval.
y Stratified by region and adjusted for town population density, density squared, and proportion of town population
aged >50 years.
Am J Epidemiol 2006;163:171–180
Prediction of West Nile Virus Using Crow Clusters
The two evaluation tools produced similar relative risk
point estimates and associated confidence intervals, which is
reassuring because it is generally expected (33–35); yet, we
are presenting a rather novel application of both methods.
Note, however, that the proportional hazards estimates had
consistently larger confidence intervals and are therefore
more conservative. Either method may be appropriate, although proportional hazards modeling has been treated as
the ‘‘gold standard’’ for comparing other methods of estimating relative risk, primarily because it requires the fewest
assumptions (33).
Aside from human population density, we cannot exclude
the possibility of other confounders we are unaware of and
did not adjust for. Covariates that are unaccounted for,
whether or not they are confounders, may also lead to residual spatial autocorrelation since we are not analyzing
a random sample of individuals but a population distributed
over space. For example, two friends from adjacent towns
may be bitten by infected mosquitoes while walking together in a local park. By simple geographic proximity, they
are not behaving independently of each other. Future research could more formally evaluate the presence of residual spatial autocorrelation; however, this is not expected to
be a limiting factor in our conclusions, and, therefore, we
did not complicate our models by adding a random effect for
spatial location.
Surveillance for dead crow clusters is limited by observation bias since it depends on public reporting. Although
we adjusted for human population density, other factors,
such as differences in public awareness among the different
regions and/or weeks, may still influence observation bias.
Future research should aim to reduce observation bias and
improve the geospatial estimation of WNV activity by incorporating covariate information about human demographics (36) along with climatic (37) and environmental (36, 38)
data. For example, Poisson regression may be used to model
counts of dead crows, or laboratory-confirmed WNV infection in any species, by using predictor variables derived
from maps of land cover/vegetation, land use, wetlands,
water bodies, temperature, previous year’s WNV activity,
and so forth. Both environmental and human demographic
covariates have been significantly associated with human
WNV infection in the Chicago, Illinois, area (36). With 6
years of data now in hand for New York State, models could
possibly be developed and calibrated for predicting WNV
activity with greater accuracy than can be obtained by using
just the current year’s dead crow clusters.
WNV continues to spread across the United States at an
epizootic level. It is the leading mosquito-borne disease in
the United States in the past century in terms of morbidity
and mortality, and it has spread northward to Canada and
southward to the Caribbean, Mexico, Central America, and
South America (39, 40). Confirmation of viral activity by
testing dead birds continues to be an excellent method of
verifying annual recurrence of risk in most areas of the
United States (41). However, with the disease capable of
rapid amplification in its bird-mosquito cycle, basing prevention and control decisions on laboratory results can be
difficult for several reasons. Unless rapid field testing kits
are used, there can be delays between identifying a dead bird
Am J Epidemiol 2006;163:171–180
179
and obtaining definitive laboratory test results. Taking action once humans have been exposed may still reduce the
total number of human infections but will be less effective
than using earlier signals of increasing viral activity. Dead
crow sightings have been demonstrated to be a valuable
crude indicator of viral activity, and the statewide evaluation
reported in this paper provides quantitative evidence of increased risk for humans living near dead crow clusters.
Thus, jurisdictions responsible for WNV surveillance
should consider cluster detection methods to identify focal
areas of viral activity sufficiently early enough to provide
personal protection warnings or to conduct mosquito control
activities.
ACKNOWLEDGMENTS
Work on this study was partially supported by federal
funds from National Science Foundation grant 9983304
entitled, ‘‘Developing a National Infectious Disease Information Infrastructure: An Experiment in West Nile Virus
and Botulism’’; from the National Institute of Allergy and
Infectious Disease, National Institutes of Health, under
contract N01-A1-25490; and from the Centers for Disease
Control and Prevention under cooperative agreement U50/
CCU223671.
The authors thank the local health departments and the
US Department of Agriculture’s Wildlife Services New
York office for WNV surveillance reports; the New York
State Department of Environmental Conservation’s Wildlife
Pathology Unit and the New York State Department of
Health (NYSDOH) Wadsworth Center’s Arbovirus Laboratory for processing and testing dead birds; the NYSDOH
Arthropod-borne Disease Program and the Wadsworth
Center’s Encephalitis PCR and Diagnostic Immunology
Laboratories for human WNV surveillance; Yoichiro
Hagiwara for managing the New York State dead bird
surveillance system; the NYSDOH Healthcom network
for system development and support; and staff and students
of the NYSDOH Zoonoses Program and Bureau of Communicable Disease Control for their contribution to the
WNV surveillance system and the dead crow maps.
The contents of this report are solely the responsibility of
the authors and do not necessarily represent the official
views of the supporting agencies.
Conflict of interest: none declared.
REFERENCES
1. Morse DL. West Nile virus—not a passing phenomenon.
N Engl J Med 2003;348:2173–4.
2. Turell MJ, Sardelis MR, Dohm DJ, et al. Potential North
American vectors of West Nile virus. Ann N Y Acad Sci
2001;951:317–24.
3. Outbreak of West Nile-like viral encephalitis—New York,
1999. MMWR Morb Mortal Wkly Rep 1999;48:845–9.
4. Update: West Nile virus activity—Eastern United States, 2000.
MMWR Morb Mortal Wkly Rep 2000;49:1044–7.
180 Johnson et al.
5. Eidson M, Kramer L, Stone W, et al. Dead bird surveillance
as an early warning system for West Nile virus. Emerg Infect
Dis 2001;7:631–5.
6. Gotham IJ, Eidson M, White DJ, et al. West Nile virus: a case
study in how New York State health information infrastructure
facilitates preparation and response to disease outbreaks.
J Public Health Manag Pract 2001;7:79–89.
7. Andrle RF, Carroll JR, eds. The atlas of breeding birds in
New York State. Ithaca, NY: Cornell University Press, 1988.
8. Eidson M, Komar N, Sorhage F, et al. Crow deaths as a sentinel
surveillance system for West Nile virus in the northeastern
United States, 1999. Emerg Infect Dis 2001;7:615–20.
9. Eidson M, Miller J, Kramer L, et al. Dead crow densities
and human cases of West Nile virus, New York State, 2000.
Emerg Infect Dis 2001;7:662–4.
10. Julian KG, Eidson M, Kipp AM, et al. Early season crow
mortality as a sentinel for West Nile virus disease in humans,
northeastern United States. Vector Borne Zoonotic Dis 2002;
2:145–55.
11. Mostashari F, Kulldorff M, Hartman JJ, et al. Dead bird
clusters as an early warning system for West Nile virus activity. Emerg Infect Dis 2003;9:641–6.
12. Theophilides CN, Ahearn SC, Grady S, et al. Identifying West
Nile virus risk areas: the dynamic continuous-area space-time
system. Am J Epidemiol 2003;157:843–54.
13. Kulldorff M. A spatial scan statistic. Communications Stat
Theory Methods 1997;26:1481–96.
14. Knox GE. The detection of space-time interactions. J R Stat
Soc (C) 1964;13:25–9.
15. Bailey TC, Gatrell AC. Interactive spatial data analysis. Essex,
United Kingdom: Longman Scientific and Technical, 1995.
16. Watson JT, Jones RC, Gibbs K, et al. Dead crow reports and
location of human West Nile virus cases, Chicago, 2002.
Emerg Infect Dis 2004;10:938–40.
17. MapInfo Corporation. MapMarker, version 8.0. Troy, NY:
MapInfo Corporation, 2002.
18. Encephalitis or meningitis, arboviral (includes California
serogroup, Eastern equine, St. Louis, Western equine, West
Nile, Powassan). Atlanta, GA: Centers for Disease Control and
Prevention. (http://www.cdc.gov/epo/dphsi/casedef/
encephalitiscurrent.htm).
19. Kulldorff M, Information Management Services, Inc.
SaTScan: software for the spatial, temporal, and space-time
scan statistics, version 6.0. 2005. (http://www.satscan.org/).
20. Recuenco S, Eidson M, Schmit KJ, et al. Cluster detection for
forecasting West Nile virus human cases. Presented at the
131st Annual Meeting of American Public Health Association,
San Francisco, California, November 15–19, 2003. (http://
apha.confex.com/apha/131am/techprogram/paper_53079.htm).
21. MapInfo Corporation. Vertical Mapper, version 3.0. User
guide. Troy, NY: MapInfo Corporation, 2003.
22. MapInfo Corporation. MapInfo Professional, version 7.5. User
guide. Troy, NY: MapInfo Corporation, 2003.
23. US Environmental Protection Agency, Office of Water.
Wetlands and West Nile virus. (EPA 843-F-04-010, August
2004). (http://www.epa.gov/owow/wetlands/pdf/WestNile_pr.
pdf).
24. Petersen LR, Marfin AA. West Nile virus: a primer for the
clinician. Ann Intern Med 2002;137:173–9.
25. McCullagh P, Nelder JA. Generalized linear models.
London, United Kingdom: Chapman and Hall, 1989.
26. Campbell GL, Marfin AM, Lanciotti RS, et al. West Nile
virus. Lancet Infect Dis 2002;2:519–29.
27. Cox DR. Regression models and life tables (with discussion).
J R Stat Soc (B) 1972;34:187–220.
28. Efron B. The efficiency of Cox’s likelihood function for
censored data. J Am Stat Assoc 1977;72:557–65.
29. Collett D. Modeling survival data in medical research. Boca
Raton, FL: Chapman and Hall/CRC, 1994.
30. Patil GP, Taillie C. Upper level set scan statistic for detecting
arbitrarily shaped hotspots. Environ Ecol Stat 2004;11:
183–97.
31. Duczmal L, Assuncao RM. A simulated annealing strategy
for the detection of arbitrarily shaped spatial clusters. Comp
Stat Data Anal 2004;45:269–86.
32. Tango T, Takahashi K. A flexibly shaped spatial scan
statistic for detecting clusters. Int J Health Geog 2005;4:11.
33. Callas PW, Pastides H, Hosmer DW. Empirical comparisons
of proportional hazards, Poisson and logistic regression modeling of occupational cohort data. Am J Ind Med 1998;33:
33–47.
34. Laird N, Olivier D. Covariance analysis of censored survival
data using log-linear analysis techniques. J Am Stat Assoc
1981;76:231–40.
35. Ma R, Krewski D, Burnett RT. Random effects Cox models:
a Poisson modeling approach. Biometrika 2003;90:157–69.
36. Ruiz MO, Tedesco C, McTighe T, et al. Environmental and
social determinants of human risk during a West Nile virus
outbreak in the greater Chicago area, 2002. Int J Health Geogr
2004;3:8.
37. Rogers DJ, Myers MF, Tucker CJ, et al. Predicting the distribution of West Nile fever in North America using satellite
sensor data. Photogram Eng Remote Sens 2002;68:112–14.
38. Shaman J, Day JF, Stieglitz M. Drought-induced amplification
of Saint Louis encephalitis virus, Florida. Emerg Infect Dis
2002;8:575–80.
39. Gould LH, Fikrig E. West Nile virus: a growing concern?
J Clin Invest 2004;113:1102–7.
40. Valero N. West Nile virus: a new challenge? Invest Clin
2003;44:175–7.
41. West Nile virus activity—United States, June 16–22, 2004.
MMWR Morb Mortal Wkly Rep 2004;53:535–6.
Am J Epidemiol 2006;163:171–180