Download as pdf or txt
Download as pdf or txt
You are on page 1of 29

Running head: ECOLOGICAL NICHE MODELING OF WEST NILE VIRUS 1

Ecological Niche Modeling of West Nile Virus Transmission

Potential in Kent County, Michigan

Kendall James Frimodig

Grand Valley State University


Running head: ECOLOGICAL NICHE MODELING OF WEST NILE VIRUS 2

Abstract

West Nile Virus (WNV) is a vector transmitted flavivirus that is transmitted through an enzootic

cycle maintained by avian and mosquito species where Humans are a dead-end host. Predictive

modeling has proven as a useful tool to inform local West Nile surveillance programs as it can

identify high risk transmission areas; however, the extent to which these models can be applied is

confined to the specific geography of the study area. This study builds upon the work of previous

research in that it outlines the predictors influencing enzootic West Nile Virus transmission

specific to a region where no such study has been conducted to date. Following a hotspot

analysis to determine clusters of human WNV infection, the geospatial differences in values of

ecological factors known to modulate enzootic transmission of West Nile were compared

between 30m x m cells in proximity to human infection and selected control cells for the years

2002, 2005, 2006, and 2012. Once significant predictors of WNV transmission were determined,

a logistic regression model was developed to produce a map where values for each cell’s

ecological variables resulted in a logit value indicating the potential for WNV transmission. In

addition to contributing to the literature the study’s findings are anticipated to direct and inform

vector collection sites of the Kent County Health Department’s vector control program, so that

the sensitivity for detecting the presence of enzootic transmission of West Nile may be enhanced.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 3

Background

West Nile Virus (WNV) is a flavivirus that is transmitted through an enzootic cycle

maintained by avian reservoirs (notably the American Robin, Crow, and Blue Jay) and a few

mosquito species of the Culex genera (tarsalis, pipiens, and quinquefasciatus) where humans are

a dead-end host. The first domestic evidence of WNV transmission appeared as a clustering of

meningoencephalitis cases in New York City circa 1999, and by the year 2001 WNV had rapidly

spread throughout the entire contiguous United States (Nash et al., 2001). In the eight years

following the first domestic WNV outbreak in 1999, approximately 29 million cases of human

WNV infection were reported from 1,869 counties within all 47 continental states except Maine

(Lindsey, Staples, Lehman, Fischer, & Centers for Disease Control and Prevention [CDC],

2010). (Krow-Lucal, Lindsey, Lehman, Fischer, & Staples, 2017). In 2015, 95% of arboviral

infections were determined to be WNV, 67% of which were classified as neuroinvasive disease

with an incidence of 0.45 cases per 100,000 population (Krow-Lucal et al., 2017). Despite this

figure, approximately 80% of infections are asymptomatic and less than 1% of infected

individuals experience manifestations of neuroinvasive disease (Campbell, Marfin, Lanciotti, &

Gubler, 2002; Hayes et al., 2005; Nash et al., 2001). Non-neuroinvasive West Nile infection

manifests in pronounced fever, muscle pain, and gastrointestinal disruptions (Campbell et al.,

2002).

Although the majority of WNV infections manifest in manageable symptoms, infection

can be far more detrimental and in certain cases lethal to vulnerable populations. Children and

the elderly are particularly vulnerable to WNV as infections can occasionally result in

neuroinvasive encephalitis; lasting neurological defects may contribute significantly to morbidity


PREDICTIVE RISK MODELING OF WEST NILE VIRUS 4

in infected children, while immunosuppression places the elderly at a higher risk of mortality

from the disease (Sejvar & Marfin, 2006). Roughly half of all reported West Nile infections are

neuroinvasive manifestations of the disease, however estimates from the CDC indicate that less

than 1% of all West Nile Infections result in neuroinvasive symptoms. The under-reporting of

less severe WNV infections results in underestimations of the true presence of enzootic WNV

transmission. Although the incidence of WNV in the United States has stabilized in recent years,

this may be partially due to successful implementation of surveillance and control practices.

Continued efforts of West Nile surveillance at the national, state, and local level are necessary to

prevent resurgent epidemics from occurring, as the disease remains endemic in most regions of

the United States (Krow, Lindsey, Lehman, Fischer, & Staples, 2017).

Predictive modeling for West Nile transmission has been proven as a useful tool to

inform local vector surveillance and control programs, as it can identify specific neighborhoods

where the potential for enzootic transmission is greater; however, the extent to which these

models can be applied is often confined to the specific geography of the study area. The lack of

homogeneity within the literature has indicated the environmental and landscape factors

associated with WNV transmission vary significantly by region and climate within the United

States (DeGroote & Sugumaran, 2012, Hartley et al., 2012; Kala, Tiwari, Mikler, & Atkinson,

2017; Larson et al., 2010; Zou, Miller, & Schmidtmann, 2007). As there currently is no uniform

model for predicting geographic regions prone to a higher risk of WNV transmission, many local

vector control programs do not have a means of quantifying where surveillance practices are

most likely to capture WNV infection in mosquito populations. The researchers personal

experience in conducting field work for the Kent County Health Departments (KCHD) WNV
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 5

surveillance and control program informs the present study, as the current mosquito collection

sites are unsupported by an evidence-based assessment and human infection rates have in recent

years exceeded the number of mosquito pools testing positive for the virus.

Literature Review

The national and state response to WNV control and prevention has been successful in

attenuating epidemics of the magnitude seen in the early 2000’s, despite the fact that endemic

transmission continues to occur across the United States. West Nile transmission occurs

primarily in the summer and fall months, when Culex mosquitos are most active. Historically

94% of human infections have been reported between the months of July and September, and

two-thirds of all cases occurred in a 6 week period from late July into the end of August (Lindsey

et al., 2010). In the two decades following the introduction of WNV to the United States most

research concerned with this issue has focused on understanding the ecological factors that may

mitigate potential outbreaks, determining the viral competence of various species and preventing

new host species from emerging, investigating shifts in climate patterns that have affected its

distribution regionally, and methods of predicting where transmission is likely to occur through

ecological modeling.

In order to use a predictive model to determine which human populations are at the

greatest risk of WNV infection, the ecological factors contributing to enzootic transmission

cycles should be contextually represented in a geospatial manner that accounts for both the

landscape features and ecological interference of the human environment; predictive modeling

for WNV is dependent upon a comprehensive understanding of the pertinent avian hosts and

their habitat, understanding the life cycle dynamics of competent vector species feeding on these
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 6

avian reservoirs, and how features of the urban human environment modulate the intersection of

these species.

Transmission Dynamics

Evidence from dead bird reporting in the WNV outbreaks occurring in the years 2002-

2006 suggests that the two predominant avian reservoirs in Kent County, Michigan are the

American Crow and Blue Jay (Kent County Health Department, n.d.). In addition to

understanding the environmental factors suitable for these avian hosts, the life cycle of the Cx.

pipiens mosquito species will be discussed in detail as this is the only known species

contributing to West Nile transmission cycles in West Michigan (Krow-Lucal et al., 2017).

Avian reservoirs. Avian hosts can survive WNV infection and develop permanent

immunity, although certain species are more susceptible to severe infection and die (Komar et

al., 2001). Resident birds that survive infection and develop lasting immunity effectively

increase the level of herd immunity among these hosts populations. Over time this may lower the

potential for enzootic transmission to occur, and this may be a key factor explaining the lowered

incidence of WNV in the United States over the last decade (Krow-Lucal et al., 2017). Despite

the observed decrease in WNV incidence in the United States, outbreaks continue to occur on a

periodic basis. One possible explanation for this periodicity is a pattern also observed with

measles infection in humans, where herd immunity eventually drops below a point due to the

death of immune cohorts and the accumulation of birth cohorts yet to be exposed to an outbreak

(Farajollahi, Fonseca, Kramer, & Marm Kilpatrick, 2011).

Vector species. Although WNV has been detected in 65 species of mosquitos in the

United States, only three species (Cx. pipiens, Cx. quinquefasciatus and Cx. tarsalis) act as
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 7

significant reservoirs for the enzootic transmission cycle. The primary vector for WNV

transmission in the study area is Culex pipiens, and as a result this study will discuss WNV

vector ecology in the specific context of this species.

Breeding dynamics. Oviposition behavior in Cx. pipiens is driven by an acute olfactory

system which senses chemical signals indicating suitable breeding sites. Blood-fed females are

attracted to potential breeding sites by several sensory cues, including organic matter and

pheromones. Crude visual sensory information also directs mosquitos, with dark and low-lying

breeding sites being attractive to ovipositioning females (Barbosa, Souto, Eiras, & Regis, 2007;

Reiskind & Wilson, 2004; Sullivan, Liu, & Syed, 2014). Urban breeding sites such as low places

with poor drainage, urban catch basins, roadside ditches, sewage treatment lagoons, and

manmade containers around houses have been associated directly with larval development of Cx.

pipiens mosquitos and WNV transmission potential in previous assessments (Hayes et al., 2005).

During ovipositioning, mosquitos are directed away from breeding sites where other

species have already laid egg rafts. This process is facilitated through the release of pheromones

specific to a species larvae; conversely larval pheromones of like species have been found to

attract ovipositioning mosquitos (Jackson, Paulson, Youngman, Scheffel, & Hawkins, 2005). As

a result of this interspecies competition, the breeding habitat of each species can often be

modeled accurately with a narrow range of ecological variables. Cx. pipiens are more adapted to

the urban environment, and larval sampling by Gardner et al. (2013) demonstrated that urban

breeding sites such as catch basins are dominated by this species; during the studies three-month

collection period, all larval specimens identified in catch basins were of the Cx. pipiens species

(Gardner et al., 2013).


PREDICTIVE RISK MODELING OF WEST NILE VIRUS 8

Ecological Factors Associated with Transmission

The breeding and feeding patterns of Cx. pipiens mosquitos are influenced by several

factors which include mean temperature and daily temperature disparity, avian reservoirs,

developmental stages of the vector, sources of artificial habitats, mosquito control practices,

aquatic ecosystems, terrestrial vegetation, and features of the natural and urban landscape

(Gardner et al., 2013; Pecoraro et al., 2007). In addition to these identified predictors of Cx.

pipiens distribution, it is also important to consider the ecological factors that facilitate avian

host habitats as this is a key component to the enzootic transmission cycle. Variables such as tree

canopy cover and impervious land cover are included in the model to account for the distribution

of resident avian populations; additionally impervious land cover is negatively associated with

vector dispersion as mosquitoes navigation is based on natural sensory cues, to such a degree that

they are unlikely to even cross a developed road (Jackson et al., 2005). Although Cx. pipiens is

known as an urban species, this instinctual navigation has been posed as an explanation as to

why areas of moderate urban development and moderate vegetation yield greater mosquito

abundance; one study investigating the effect of differing urban classes within the cities of

Chicago and Detroit found that WNV transmission was greatest among inner suburbs with 1940-

1960 era housing, moderate vegetation cover, and moderate population density (Ruiz, Walker,

Foster, Haramis, & Kitron, 2007).

Urban environment. Socio-economic factors such as per-capita income and housing

quality have also been demonstrated to increase the risk of West Nile transmission (Ruiz et al.,

2007). The authors suggest that the relationship with socioeconomic status may potentially due

to the fact that lower income neighborhoods are older and less attention is directed to
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 9

infrastructure management within these neighborhoods, specifically unmaintained drainage

systems which produce urban habitats for vector populations (Pecoraro et al., 2007; Ruiz et al.,

2007). Other studies have suggested that urban containers in yards of low income neighborhoods

are more abundant and maintained less than neighborhoods of higher income. Additionally,

lower income neighborhoods are more likely to have pre-1970’s housing, which has been

associated with the clustering of WNV infection among humans.

Study Designs

Humans do not act as a natural reservoir for the transmission of WNV, as the viral load in

infected individuals is not sufficient to transmit the disease to mosquitos (Hayes et al., 2005). As

humans are not a competent natural host within this transmission cycle, prior studies of WNV

transmission have predominantly focused on the geospatial analysis of environmental and urban

landscape factors that result in the convergence of avian host reservoirs and known vector

species, which together indicate transmission prone areas. These geographical regions prone to

WNV transmission are represented by a surrogate indicator such as human infection as data for

both avian and mosquito infection rates are rarely available together, while the availability of

either is highly dependent on the resources of the local surveillance and control program (Centers

for Disease Control and Prevention [CDC], 2013).

Predictive modeling. In light of the apparent the local and regional differences

influencing WNV transmission risk, several predictive modeling assessments have started with a

large base of ecologic factors that have been identified by prior investigations. Prior modeling

assessments have differed in the process of variable selection and the geospatial parameters

assigned to these variables, in addition to the outcome representing WNV transmission used. A
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 10

common finding among the literature is the variation in the factors that are ultimately included in

the final predictive model. The ecological and landscape factors determined to be significant by

prior modeling assessments are summarized in Appendix A1. Several of these assessments have

taken place in the city of Chicago, while one assessment was also conducted in Detroit, MI; the

identified ecological variables of importance to WNV transmission by these studies will be

included in the present study as both cities lie within the same NOAA climatic region. These

variables include housing age, urban land cover, and tree canopy density. Additional variables

chosen for this study include elevation, the minimum, maximum and average slope, distance to

wetlands, streams, major roadways, and population density as well as housing age.

Spatial considerations. The reliability of human WNV infection in using patient

addresses and dispersion of relevant avian and vector species has been considered and accounted

for in the majority of previous modeling studies. Several methods have been utilized to account

for the effect of spatial interaction; inverse distance weighting of ecological values accounts for

the effect of neighboring landscapes, and most studies have found that these spatial interactions

are negligible beyond the 1-2 km area surrounding. In using human infection data, the location of

infection is determined to be the infected individuals address, which is not necessarily the site of

virus transmission. Despite this uncertainty, when multiple cases of WNV in humans occur in the

same neighborhood, it is more likely the neighborhood is a ‘hot spot’ for viral transmission and

thus the random dispersion of human movement has not been a primary consideration in most

WNV research. In an assessment of the environmental predictors for Cx. pipiens abundance,

Trawinski & Mackay (2010) found optimal significance at buffer sizes of 1km. These findings
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 11

support the notion that the relationship of neighborhood level ecological features and Cx. Pipiens

abundance is not likely confounded by selecting a relatively small geographic buffer.

Study objectives. As the WNV surveillance and control program of the KCHD has a low

capacity for vector collection with just 10 gravid traps, early indication of enzootic transmission

is dependent on the placing of these traps in habitats suitable for both avian and vector hosts

involved in the transmission cycle. This study aims to identify the possible existence of

geospatial differences in ecologic and landscape factors occurring within neighborhood level

clusters of human WNV infection and comparable urban areas where cases did not occur, for the

outbreak years of 2002, 2005 and 2006.

Methods

Approval of Research

This study was determined to be non-human subjects research by the Human Research

Review Committee at Grand Valley State University. The original human WNV case data

consisted of an ID, residential address of the patient, and year however this data was de-

identified prior to the researcher’s acquisition from the KCHD; residential addresses for human

WNV cases were first geocoded by personnel at the KCHD, and were then converted to a

ArcMap shape file of a 50m x 50m resolution so that the case households could not be

potentially identified by the researcher, however the approximate geographical locations of

infection were retained.

Data Collection

Data for the variables included in this analysis were retrieved from publicly available

satellite imagery databases such as USGS, in addition to the environmental health and infectious
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 12

disease division at the KCHD. Four separate shape files for human WNV cases were collected

for the years 2002, 2005, 2006, and 2012, in addition to an aggregate shape file for all of these

high transmission years. Environmental data was obtained for the median year of 2006; remote

sensing data for canopy cover, imperviousness, and elevation at 30x30 meter resolution for all

years was obtained from the from the Multi-Resolution Land Characteristics Consortium 2011

revision data sets for national land cover (Multi-Resolution Land Characteristics Consortium

(MRLC), 2011a; MRLC, 2011b.). Significant differences were not found between these raster

values for the 2001 and 2011 datasets. Shapefiles for parcel dispersion, catch basin and retention

pond location, streams, parks, and developed roads were obtained from the division of

environmental health at the KCHD. A Normalized Difference Vegetation Index (NDVI) for the

study area was derived from the National Agricultural Imagery Project high resolution

orthoimagery, at a 1x1 meter resolution. Data for housing age, population density, and area of

water were derived from census block group data sourced from the U.S. Census Bureau.

Data protection and use. Throughout the study, the shapefiles obtained from the KCHD

were kept on an encrypted external hard drive at the investigators personal residence. The

researchers machine was password protected in order to prevent local access. The folder residing

on the encrypted external hard drive was password protected using Folder Lock (v.7) so that the

folder was only visible during its use.

Software used. Spatial implementation of ecological variables and human case layers

were conducted using ArcMap (v10.5). Clustering of human infection were determined using the

Spatial Analyst toolset within ArcMap, in addition to zones of low clustering within the urban

environment of Kent County, MI. The attribute table for areas of low clustering produced by
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 13

ArcMap was imported into SAS (v9.4) where control cells were randomly selected from within

these clusters in order to yield a comparable number of control cells for statistical analysis.

Preparation of Predictor Variables. Ecological and landscape variables were obtained

for the years 2001, 2006 and 2012. In order to determine the change in these factors over the

study period of 2002-2006, pairwise T-tests were conducted for each cells value between the

years 2001 and 2006. Statistical differences were not found between the years and data obtained

for the year 2006 were used for the remainder of the study. For each independent variable

separate raster grids were produced using IDW at varying distance bands to account for

neighborhood characteristics; interpolated raster files at 0.25, 0.5, 1.0 and 1.5km buffers were

produced.

Case-control preparation. A case-control design was chosen with the primary unity of

measurement being a binary infection outcome among any parcel within the study area. The

exclusion of rural areas of Kent County was appropriate as the only competent vector species Cx.

Pipiens has been shown to be a predominantly urban species; Larson et al. (2010) found that

89% of Cx. Pipiens samples fell inside or within 4km of an incorporated city boundary.

Additionally, only 3 of the final 116 human infections for all years were found to occur outside

the urban landscape of Kent County. The final study area encompassed a rectangular boundary

including the entire urban landscape of Kent County with 14 of 30 total civil divisions. A

30x30m sampling grid was overlaid on the study area, and cells falling outside parcel blocks

were removed.

Hotspot analysis. The outlier analysis and hotspot analysis tools were used in ArcMap to

determine potential cases to be left out of the analysis. All viable cases within the study area
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 14

demonstrated clustering at a 2.0km zone of indifference. Analysis at 750m, 1000m, and 1500m

threshold distances were also conducted; these analyses indicated fewer and more localized

clusters, and a threshold range of 1000m was selected in order to account for both the average

(200-1000m) and maximum (2000m) known dispersion ranges of Cx. pipiens. In using a zone of

indifference for the cluster analysis, the value for each cell representing a possible ‘household’

was a function of every other cell in the study area; however, the weight of cells within a 1000m

range were given equal weight, and those outside the distance were inversely weighted with a

cubic factor. Each cell was assigned a value of its local z-score representing local clustering.

Case selection. All case clusters occurred within the urban environment and thus a

geographical boundary ruling out 3 WNV human cases occurring in rural areas was utilized. The

resulting case sample included 113 cells of the total 174,546 possible parcel cells.

Control selection. In addition to determining geospatial hotspots for WNV infection, the

spatial analyst tool in ArcMap was also used to determine control areas or ‘coldspots’ which

accounted for the spatial arrangement of global clustering through the zone of indifference

function. Clustering data was imported into SAS and 500 cells were randomly selected from the

9000 cells in the lowest quantile of standardized z-scores.

Statistical analysis. Attribute tables for the case-control data were converted into

shapefiles in ArcMap, and the sampling tool as used to obtain raster values for each ecologic

variable at the specified cell location. The final case-control attribute table was imported into

SAS for analysis. A logistic model was developed in order to assess whether significant

differences in ecological and landscape variables exist between case and control locations. The

presence of a human WNV case was modeled as indirect binary outcome representing the
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 15

presence of WNV transmission. Values for each landscape or ecological variable at all buffer

ranges retained were included as separate independent variables. These interpolated variables

included canopy cover, imperviousness, NDVI, catch basin density, and road density.

Logistic model application. Once the final logistic model was developed, each 30m cell

within the entire county was assigned a logit value. This logit value was derived from

multiplying each cell’s ecological values by the logistic model parameter estimates, which were

then modeled as a probability from 0 to 1 through an inverse logistic transformation of the model

output.

Results

Hotspot analysis were conducted on urban and suburban parcels within a 30x40 kilometer

area of Kent County. Clustering was observed for WNV infections for years 2002-2012, even

with the exclusion of nearly 50% of the county area; this is likely due to the habitat preference of

Cx. pipiens, although population density was included in the model for adjustment of this factor.

The results of the hotspot analysis at 4 different threshold distances are depicted in Figure 1.

Figure 1. Hotspot analysis of WNV confirmed cases at 1000, 1250, 1500, and 2000-meter buffers

1000m 1250m 1500m 2000m

Note. Color ranges represent 15 category quantiles of local z-scores for each 30x30 cell withn residential areas.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 16

Although the hotspot analysis threshold distances demonstrated similar central clustering, the

distance band of 1000m showed a greater cluster tolerance. This allowed for control selection

within the urban environment where infections were not present for all outbreak years.

Interestingly, not a single case occurred for the 10-year period in the townships of Cannon,

Cascade, and Ada. Although case cells we’re proportionally more in heavily urbanized

neighborhoods and control cells were more densely located in suburban neighborhoods, there

was not a complete segregation within the sample. Figure 2 below illustrates the case and control

locations, as well as percent pre-1970’s housing and major roadways. These two variables were

found to have a strong correlation with the case and control clusters.

Figure 2. Percent pre-1970’s housing by census group, major


roadways, and case and control cells.

% Pre-1970’s housing

0-10
10-20
20-30
30-40
40-50
50-60
60-70
70-80
80-90
90-100
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 17

In addition to apparent differences in housing age between case and control cells, several

other variables were found to differ in the logistic regression analysis. The mean values of each

variable included in the final logistic regression model are provided for case and control cells in

Table 1 below.

Table 1.
Characteristics of socioeconomic and ecological variables at case and control cells

Case Cells (n=113) Control Cells (n=500)

Variable Mean SD Mean SD

% 1940-1970’s Housing 65.16 29.940 23.91 16.073

% Pre-1940’s Housing 27.69 28.692 3.38 4.116

% Imperviousness 1.0km 43.89 14.902 19.42 13.0157

Population Density 2520.59 2016.46 752.85 676.399

Distance to Wetlands (m) 681.41 537.336 242.88 194.421

Distance to Detention Basin (m) 1024.74 585.632 841.14 759.852

Distance to Major Roadway (m) 159.97 142.697 385.86 310.388

Distance to Streams (m) 570.16 425.728 296.65 234.608

Water Density 30771.75 134687.75 225166.08 353363.39

Overall case cells were closer to major roadways, had higher impervious land cover, higher

population density and percent pre-70’s as well as pre-40’s housing, and had lower water

density. Significant differences were identified between case and control cells for all of the

variables included in the final logistic regression model, with the exception of distance to

detention basins. This variable was included as it increased the Harrell’s C statistic by 0.3

percent.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 18

Logistic Regression.

The odds of West Nile Virus infection increased with further distances to detention

basins and wetlands, increased water density, population density, impervious land cover, and

percent 1940-1970’s housing as well as pre-1940’s housing. The odds of human infection

decreased with further distances to the nearest major roadway or stream. Although distances to

streams were greater in Table 1. for case cells, the standard deviation was larger than the mean

distance for control cells. The final model predicted the correct probability of the 613 cells with

93% concordance. The strongest predictor of WNV human infections was the percent pre-40’s

housing (OR=1.13, p=<0.001). This is hypothesized to be associated with the infrastructure of

these neighborhoods being older and less maintained, creating the potential for standing water in

aged storm water infrastructure.

Table 2.
Final logistic regression parameter estimates, modeled with infection as the outcome.

B S.E. Sig. Exp(B)

Intercept -4.33 0.617 <0.001 0.01

% 1940-1970’s Housing 0.0181 0.008 0.028 1.02

% Pre-1940’s Housing 0.1180 0.025 <0.001 1.13

% Imperviousness 1.0km 0.0369 0.0142 0.009 1.04

Population Density 0.000212 0.000212 <0.001 1.00

Distance to Wetlands 0.00146 0.000795 0.065 1.16*

Distance to Detention Basin 0.000358 0.000318 0.259 1.04*


PREDICTIVE RISK MODELING OF WEST NILE VIRUS 19

Distance to Major Roadway -0.00207 0.000995 0.037 0.82*

Distance to Streams -0.00189 0.000778 0.015 0.83*

Water Density 2.57E-6 1.24E-6 0.038 1.00

Notes.
* Parameters are interpreted as a 100m increase in the distance to the nearest feature

Mapping Transmission Risk

Of the 613 cells incorporated in the logistic regression, the final model was able to

predict the presence or absence of infection with 93% concordance (Harrell’s c=0.930). With the

predictive sensitivity of the model established, it was applied to 1,245,024 30x30 meter cells to

produce a final transmission risk map which is shown on the right. The resulting risk map was

not truly continuous due to population density and housing age being sampled at a census block

group level. Additionally, the parameter estimates associated with each resulted in sharp changes

at a few sub-urban census block-group boundaries. Despite this, these abrupt changes were more

than often between adjacent census block-groups where transmission risk was already quite low.

These artifacts can be observed with rural block-groups where the factors predicting risk are few,

but older housing age results in an overestimated probability of infection; however, the predicted

probabilities for these older townships are still far lower than urban areas with similar housing

age characteristics. The risk map is shown in Figure 3, and an enlarged map can be found in

Appendix B.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 20

Figure 3. Final probability map for West Nile Virus transmission risk.

Discussion

Limitations

Previous assessments have differed in their outcome used for representing WNV

transmission risk, although most recent predictive modeling studies have used human infection

as an indicator. In avian surveillance birds are rarely tested uniformly, but rather a few samples

are tested in order to indicate the presence of WNV transmission prior to human infection. Using

calculated minimum infection rates in pools of collected mosquitos is regarded as the most

robust outcome of quantifying WNV transmission, however this requires a well-equipped and

sensitive local surveillance program with rich mosquito collection data; Ruiz et al. (2010); used

mosquito infection rates as an outcome for the model however the samples were obtained from

370 unique trapping locations (Ruiz et al., 2010). Although the Kent County Health Department

collects mosquito specimens as part of its WNV surveillance and control program, specimens are
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 21

collected from just 10 sites and as a result the sample sizes of infected mosquito pools is far too

low for use in the current assessment; in the last 5 years of surveillance positive mosquito pools

for WNV were either absent or lower than 10 for each transmission season.

As the original metric of possible human infection of WNV was residential address,

hotspot analysis were only conducted on cells overlying parcel blocks. As the final outcome

being modeled was areas of higher transmission risk, cells lying within county parks could likely

be viable ‘hotspots’ and similarly highly impervious city centers could be ‘coldspots’; despite

this the hotspot analysis was conducted exclusively on parcel groups as this type of analysis

should only be inclusive of cells where a possible human infection could be recorded.

A number of prior investigations have accounted for acute and long term climatic trends

in their modeling of WNV transmission risk. Precipitation intensity and periodicity in particular

has been suggested to be a key mediator of standing water availability in urban drainage systems.

Due to a low sample size and just 4 epidemic years, temporal trends comparing between year

climatic anomalies were left out of the study. Significant variations in precipitation patterns

occurring between the outbreak years 2002-2006 could confound these studies findings as human

WNV cases occurring in 2002, 2005, and 2006 were aggregated, which consequently ruled out

any temporality. Previous investigators have used time-aggregated human infection data as an

outcome, although these studies largely were concerned with ecological niche modeling where

factors such as vegetation, storm drainage, and housing characteristics are assumed to remain

relatively static between years (Pecoraro et al., 2007; Rochlin, Turbow, Gomez, Ninivaggi, &

Campbell, 2011; Ruiz et al., 2010).


PREDICTIVE RISK MODELING OF WEST NILE VIRUS 22

As the study used ecologic data respective to the study years, the relevance of the

produced risk map for prospective WNV surveillance may not reflect the current state of the

urban environment within Kent County; housing renewal, infrastructure replacement and

maintenance, urban development and decreased vegetation cover occurring between the years

2002-2017 could change the distribution of ecological factors influencing WNV transmission

and consequently affect the generalizability of the risk map.

Although the model demonstrated a high level of concordance at 93%, this validation was

conducted upon the data which the logistic model was informed by. As a large number of the

control cells randomly selected were from sub-urban neighborhoods, the model may be biased

due to the large differences in factors such as imperviousness and housing age that characterize

urban-suburban differences rather than neighborhoods at a greater or lower risk for WNV

transmission. For example, imperviousness is known to have a bimodal association with WNV

risk where extremely high values and extremely low values are negatively associated with a

decreased risk of transmission.

An exclusive urban analysis would likely indicate a negative relationship with

imperviousness, whereas the current study suggests a positive relationship due to the inclusion of

sub-urban neighborhoods; this correctly identifies the risk difference between case and control

cells inclusively but when the model was applied to the entire study area, the risk was likely

overestimated for highly impervious areas where no cases were observed to occur. As the study

included all cases, there were few urban controls to be selected outside a 2.0km range of an

infection. Had the case cells outside of hotspots been excluded a wider range of controls could

likely be chosen; as residential addresses are not an exact measure of where infection took place,
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 23

it is likely that a number of outlier cases were not infected at their residence. Hot-spot analysis

can determine these outliers; however, it is also possible that these cluster analyses can be

informed by cases falling to the same fallacy. Due to this uncertainty, all confirmed cases were

included.

Although a large portion of the available human infection data occurring in the year 2012

was originally reserved for cross validation of the model so that its strength over time could be

assessed, an aggregation of the data for all years in the 10-year period was used instead. Cases

occurring in 2012 demonstrated comparable clustering for the years 2002-2006, there was one

exception in that a cluster of 12 cases from this year filled in a possible cold-spot if data from

2002-2006 were to be used exclusively.

Conclusion

In addition to contributing to the set of identified predictors of WNV transmission within

the literature, this study elucidates the predictors of transmission specific to the local ecology of

Kent County, Michigan where no such investigation has been conducted to date. The assessment

is also unique in the fact that the hotspot analysis was conducted exclusively on parcel spatial

relationships. Rochlin et al. (2011) utilized parcel data for control selection, however controls

were selected from any area outside the perimeters of a hotspot whereas this study assessed

outlier clustering for control selection. Given the findings of this study, the vector control and

surveillance program of KCHD could modify their collection sites accordingly and potentially

improve the sensitivity of detecting WNV presence in mosquito populations prior to subsequent

infection in human populations. An overlay of the current collection sites with infection

clustering and the final risk map are included in the appendix for reference.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 24

References

Barbosa, R. M. R., Souto, A., Eiras, Á. E., & Regis, L. N. (2007). Laboratory and field

evaluation of an oviposition trap for Culex quinquefasciatus (Diptera: Culicidae). Memorias

Do Instituto Oswaldo Cruz, 102(4), 523–529. https://doi.org/Doi 10.1590/S0074-

02762007005000058

Campbell, G. L., Marfin, A. a, Lanciotti, R. S., & Gubler, D. J. (2002). West Nile virus. The

Lancet. Infectious Diseases, 2(9), 519–529. https://doi.org/10.1016/S1473-3099(02)00368-7

Centers for Disease Control and Prevention. (2013). West Nile Virus and Other Arboviral

Diseases — United States, 2012. Mmwr, 62(25), 513–7.

DeGroote, J. P., & Sugumaran, R. (2012). National and regional associations between human

West Nile virus incidence and demographic, landscape, and land use conditions in the

coterminous United States. Vector Borne and Zoonotic Diseases (Larchmont, N.Y.), 12(8),

657–665. https://doi.org/10.1089/vbz.2011.0786

Farajollahi, A., Fonseca, D. M., Kramer, L. D., & Marm Kilpatrick, A. (2011). “Bird biting”

mosquitoes and human disease: a review of the role of Culex pipiens complex mosquitoes

in epidemiology. Infection, Genetics and Evolution : Journal of Molecular Epidemiology

and Evolutionary Genetics in Infectious Diseases, 11(7), 1577–1585.

https://doi.org/10.1016/j.meegid.2011.08.013

Gardner, A. M., Anderson, T. K., Hamer, G. L., Johnson, D. E., Varela, K. E., Walker, E. D., &

Ruiz, M. O. (2013). Terrestrial vegetation and aquatic chemistry influence larval mosquito

abundance in catch basins, Chicago, USA. Parasites & Vectors, 6(1), 9.

https://doi.org/10.1186/1756-3305-6-9
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 25

Hartley, D. M., Barker, C. M., Le Menach, A., Niu, T., Gaff, H. D., & Reisen, W. K. (2012).

Effects of temperature on emergence and seasonality of West Nile virus in California.

American Journal of Tropical Medicine and Hygiene, 86(5), 884–894.

https://doi.org/10.4269/ajtmh.2012.11-0342

Hayes, E. B., Komar, N., Nasci, R. S., Montgomery, S. P., O’Leary, D. R., & Campbell, G. L.

(2005). Epidemiology and transmission dynamics of West Nile virus disease. Emerging

Infectious Diseases, 11(8), 1167–1173. https://doi.org/10.3201/eid1108.050289a

Jackson, B. T., Paulson, S. L., Youngman, R. R., Scheffel, S. L., & Hawkins, B. (2005).

Oviposition preferences of Culex restuans and Culex pipiens (Diptera: Culicidae) for

selected infusions in oviposition traps and gravid traps. Journal of the American Mosquito

Control Association, 21(4), 360–365. https://doi.org/10.2987/8756-

971X(2006)21[360:OPOCRA]2.0.CO;2

Kala, A. K., Tiwari, C., Mikler, A. R., & Atkinson, S. F. (2017). A comparison of least squares

regression and geographically weighted regression modeling of West Nile virus risk based

on environmental parameters. PeerJ, 5, e3070. https://doi.org/10.7717/peerj.3070

Kent County Health Department. (n.d.). Kent County Map of Reported Dead Birds, 2002.

Retrieved March 11, 2018, from

https://www.accesskent.com/Health/CommDisease/pdfs/archive/deadbirds.pdf

Komar, N., Panella, N. A., Burns, J. E., Dusza, S. W., Mascarenhas, T. M., & Talbot, T. O.

(2001). Serologic evidence for West Nile virus infection in birds in the New York City

vicinity during an outbreak in 1999. Emerging Infectious Diseases, 7(4), 621–625.

https://doi.org/10.3201/eid0704.010403
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 26

Krow-Lucal, E., Lindsey, N. P., Lehman, J., Fischer, M., & Staples, J. E. (2017). West Nile

Virus and Other Nationally Notifiable Arboviral Diseases — United States, 2015. MMWR.

Morbidity and Mortality Weekly Report, 66(2), 51–55.

https://doi.org/10.15585/mmwr.mm6602a3

Larson, S. R., DeGroote, J. P., Bartholomay, L. C., & Sugumaran, R. (2010). Ecological niche

modeling of potential West Nile virus vector mosquito species in Iowa. Journal of Insect

Science (Online), 10(110), 110. https://doi.org/10.1673/031.010.11001

Lindsey, N. P., Staples, J. E., Lehman, J. A., Fischer, M., & Centers for Disease Control and

Prevention (CDC). (2010). Surveillance for human West Nile virus disease - United States,

1999-2008. Morbidity and Mortality Weekly Report. Surveillance Summaries (Washington,

D.C. : 2002), 59(2), 1–17. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/20360671

Nash, D., Mostashari, F., Fine, A., Miller, J., O’Leary, D., Murray, K., … Masatoshi, N. (2001).

The Outbreak of West Nile Virus Infection in the New York City Area in 1999. New

England Journal of Medicine, 344(24), 1807–1814.

https://doi.org/10.1056/NEJM200106143442401

Pecoraro, H. L., Day, H. L., Reineke, R., Stevens, N., Withey, J. C., Marzluff, J. M., & Meschke,

J. S. (2007). Climatic and landscape correlates for potential West Nile virus mosquito

vectors in the Seattle region. Journal of Vector Ecology : Journal of the Society for Vector

Ecology, 32(1), 22–28. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/17633422

Reiskind, M. H., & Wilson, M. L. (2004). Culex restuans (Diptera: Culicidae) oviposition

behavior determined by larval habitat quality and quantity in southeastern Michigan.

Journal of Medical Entomology, 41(2), 179–186. https://doi.org/10.1603/0022-2585-


PREDICTIVE RISK MODELING OF WEST NILE VIRUS 27

41.2.179

Rochlin, I., Turbow, D., Gomez, F., Ninivaggi, D. V., & Campbell, S. R. (2011). Predictive

mapping of human risk for West Nile virus (WNV) based on environmental and

socioeconomic factors. PloS One, 6(8), e23280.

https://doi.org/10.1371/journal.pone.0023280

Ruiz, M. O., Chaves, L. F., Hamer, G. L., Sun, T., Brown, W. M., Walker, E. D., … Kitron, U.

D. (2010). Local impact of temperature and precipitation on West Nile virus infection in

Culex species mosquitoes in northeast Illinois, USA. Parasites & Vectors, 3(1), 19.

https://doi.org/10.1186/1756-3305-3-19

Ruiz, M. O., Walker, E. D., Foster, E. S., Haramis, L. D., & Kitron, U. D. (2007). Association of

West Nile virus illness and urban landscapes in Chicago and Detroit. International Journal

of Health Geographics, 6(1), 10. https://doi.org/10.1186/1476-072X-6-10

Sejvar, J. J., & Marfin, A. A. (2006). Manifestations of West Nile neuroinvasive disease.

Reviews in Medical Virology, 16(4), 209–224. https://doi.org/10.1002/rmv.501

Sullivan, G. A., Liu, C., & Syed, Z. (2014). Oviposition signals and their neuroethological

correlates in the Culex pipiens complex. Infection, Genetics and Evolution : Journal of

Molecular Epidemiology and Evolutionary Genetics in Infectious Diseases, 28, 735–743.

https://doi.org/10.1016/j.meegid.2014.10.007

Zou, L., Miller, S. N., & Schmidtmann, E. T. (2007). A GIS tool to estimate West Nile virus risk

based on a degree-day model. Environmental Monitoring and Assessment, 129(1–3), 413–

420. https://doi.org/10.1007/s10661-006-9373-8
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 28

Appendices

Appendix A.
Identified predictors and outcomes utilized in prior predictive modeling of West Nile Virus risk

Investigator Study area Significant ecological factors identified Outcome used

Positive association with urban landscape class of Inner Suburbs,


Chicago, IL and Human WNV
Ruiz (2007) characterized by post WWII housing, moderate vegetation cover,
Detroit, MI cases
moderate population density

Positive association with urban land use and negative association


Brown et al. Northeastern Human WNV
with % forest land use, highest odds ratios observed where urban
(2008) United States cases
land use > 15.13% and forest land use < 38.29% (4.38, 4.40).

Smaller retention ponds and catch basins were associated with


Irwin et al. Mosquito
Madison, WI far greater larval development of Culex species than natural
(2008) abundance
ponds or larger retention ponds.

Positive associations with income below poverty line, total house Infected
Lie & Weng,
Chicago, IL income, and aggregate score of urban cover, water, and grass mosquito and
(2009)
density avian species

Larson et al., Aggregate of slope (steepness of terrain), aspect (direction slope Human WNV
Iowa
(2010) faces), distance to nearest urban area cases

Positive associations with urban cover, housing age, average


elevation, average and maximum slope, distance to both
Trawinski & Mosquito
Erie County, NY wetlands and hydrography. Negative associations with tree
Mackay (2010) abundance
cover, wetland density, vacant and agricultural land, and NDVI.
Variables were highly significant at a 1km buffer range.

Positive association were found with aquatic pH, ammonia, and


Gardner et al. nitrate levels as well as surrounding density of terrestrial Cx. pipiens
Chicago, IL
(2013) deciduous shrubs < 1m, and tree cover. Negative associations abundance
were found with nearby flowering shrub density.
PREDICTIVE RISK MODELING OF WEST NILE VIRUS 29

Ruiz et al. Mosquito


Midwest, Illinois
(2010) infection rates

Appendix B. Enlarged transmission risk map.

You might also like