Applying a Bayesian weighted surveillance approach to detect chronic

|

Received: 2 October 2017    Accepted: 26 March 2018 DOI: 10.1111/1365-2664.13178

RESEARCH ARTICLE

Applying a Bayesian weighted surveillance approach to detect chronic wasting disease in white-­tailed deer Christopher S. Jennelle1

 | Daniel P. Walsh2

 | Michael D. Samuel3 | Erik E. Osnas4 | 

Robert Rolley5 | Julia Langenberg5 | Jenny G. Powers6 | Ryan J. Monello6 |  E. David Demarest7 | Rolf Gubler7 | Dennis M. Heisey2 1 Minnesota Department of Natural Resources, Wildlife Health Program, Forest Lake, Minnesota; 2US Geological Survey, National Wildlife Health Centre, Madison, Wisconsin; 3US Geological Survey, Wisconsin Cooperative Wildlife Research Unit, University of Wisconsin, Madison, Wisconsin; 4US Fish and Wildlife Service, Division of Migratory Bird Management, Anchorage, Alaska; 5Wisconsin Department of Natural Resources, Madison, Wisconsin; 6Biological Resources Division, National Park Service, Fort Collins, Colorado and 7Shenandoah National Park, Luray, Virginia

Correspondence Christopher S. Jennelle, Minnesota Department of Natural Resources, Wildlife Health Program, 5463 West Broadway Avenue, Forest Lake, MN 55025. Email: [email protected] Present address Ryan J. Monello, National Park Service, Inventory and Monitoring Program, Pacific Island Network, PO Box 52, Hawaiʽi Volcanoes National Park, Hawaii 96718 Funding information Natural Resources Preservation Program Handling Editor: Silke Bauer

Abstract 1. Surveillance is critical for early detection of emerging and re-emerging infectious diseases. Weighted surveillance leverages heterogeneity in infection risk to increase sampling efficiency. 2. Here, we apply a Bayesian approach to estimate weights for 16 surveillance classes of white-tailed deer in Wisconsin, USA, relative to hunter-harvested yearling males. We used these weights to conduct a surveillance programme for detecting chronic wasting disease (CWD) in white-tailed deer at Shenandoah National Park (SHEN) in Virginia, USA. 3. Generally, for surveillance, risk of infection increased with age and was greater in males. Clinical suspect deer had the highest risk, with weight estimates of 33.33 and 9.09 for community-reported and hunter-reported suspect deer, respectively. Fawns had the lowest risk with an estimated weight of 0.001. 4. We used surveillance weights for Wisconsin deer to determine sampling effort required to detect a CWD-positive case in SHEN if prevalence in yearling males ≥0.025. The sampling required to detect CWD was 37–91 adult deer, depending on the adult male:female ratio in the surveillance stream. We collected rectal biopsies from 49 female and 21 male adult deer, and 10 additional samples from vehicle-killed deer. CWD was not detected and we concluded with 95% probability that prevalence in the reference population (yearling males) was between 0.0% and 3.6%. 5. Synthesis and applications. Our approach allows managers to estimate relative surveillance weights for different host classes and quantify limits of disease detection in real time when only a sample of animals from a population can be tested, resulting in considerable cost savings for agencies performing wildlife disease detection surveillance. Additionally, it provides a rigorous means of estimating prevalence limits when a disease/pathogen is not detected in a sample set. It is

Michael D. Samuel, Robert Rolley, Julia Langenberg, and Dennis M. Heisey are retired. This article has been contributor to by US Government employees and their work is in the public domain in the USA.

J Appl Ecol. 2018;1–10.

wileyonlinelibrary.com/journal/jpe   © 2018 The Authors. Journal of Applied Ecology |  1 © 2018 British Ecological Society

|

Journal of Applied Ecology 2      

JENNELLE et al.

therefore applicable to other wildlife, domestic animal and human disease systems, which can be characterized by surveillance classes with heterogeneous probability of infection. This methodology is also extendable to other disciplines such as invasive species, environmental toxicology, and generally, any ecological question seeking to efficiently use scarce financial and human resources to maximize the detection probability of a rare event. KEYWORDS

Bayesian, cervid, chronic wasting disease, Odocoileus virginianus, prevalence, prion, surveillance, white-tailed deer, wildlife diseases

1 |  I NTRO D U C TI O N

weights permitted managers to target surveillance efforts on demographic groups with high risk of infection. However, Walsh and Miller

With the increasing emergence and importance of wildlife diseases

(2010) utilized a frequentist statistical approach that did not account

and decreasing resources for detection and management, wildlife

for uncertainty associated with estimating class-­specific relative

managers need robust and efficient tools to detect, monitor, and

CWD risk. Failure to account for this uncertainty produces an overly

control epizootics. The early detection of diseases within wildlife

optimistic assessment of disease freedom (expectation of a smaller

populations is critical to maximize the success of management in-

number of samples to achieve surveillance goals than needed). To

terventions and reduce potential risk to human health from zoonotic

address this deficiency, Heisey et al. (2014) formulated a Bayesian

diseases. Surveillance methods and protocols for disease detection

procedure that utilized training data to estimate class-­specific rela-

in domestic livestock (Cameron & Baldock, 1998a, 1998b; Cannon

tive risks of CWD infection, which accounted for uncertainty in esti-

& Roe, 1982; Ziller, Selhorst, Teuffert, Kramer, & Schlüter, 2002)

mated weights. They also demonstrated how to use relative disease

have been available for decades (Salman, 2003). The primary goal

risk with surveillance data, where no infected individuals are found,

of wildlife disease detection surveillance is to establish whether a

to estimate population-­level CWD detection thresholds for various

region or population is “free from disease” by maximizing the prob-

demographic classes. The result was a surveillance method that used

ability of detecting a specific disease or causative agent within an

individuals from high risk disease classes, requiring fewer overall

animal population. However, the concept of freedom from disease

samples to reach an a priori detection threshold (e.g., 1% prevalence),

is a misnomer (Heisey, Jennelle, Russell, & Walsh, 2014) because it

that also accounts for class-­specific relative risk uncertainty.

is impossible in practice to prove an infected individual(s) is absent

Despite the benefits of using these weighted surveillance

without testing the entire population. Rather, this terminology im-

techniques, examples of their successful applications are limited.

plies if a disease is present, it is prevalent at or below some prespec-

In hopes of making these methods readily available to ecologists,

ified threshold deemed relevant by stakeholders and policy makers.

we utilized the (Heisey et al., 2014) Bayesian approach to estimate

Statistical methods to demonstrate that a region or population is,

relative CWD weights in white-­t ailed deer (Odocoileus virginianus)

for practical purposes, free from disease were based on this prem-

within Wisconsin, and used them to demonstrate the application

ise (Cameron & Baldock, 1998b; Cannon & Roe, 1982; OIE – World

of this type of surveillance programme for CWD in Shenandoah

Organisation for Animal Health 2011; Salman, 2003).

National Park (SHEN), Virginia, USA. We addressed four primary

While the classic approaches to disease surveillance rely on sim-

objectives: (a) develop a CWD risk assessment of white-­t ailed deer

ple random sampling (Heisey et al., 2014), recent advances can in-

in southwestern Wisconsin through the application of the Bayesian

crease sampling efficiency (i.e., requiring fewer samples to achieve

approach for estimating surveillance weights; (b) apply our resulting

surveillance goals) by employing improved sampling designs, using

Wisconsin CWD weights to a surveillance programme in SHEN; (c)

alternative data sources, and accounting for heterogeneous disease

adapt the (Heisey et al., 2014) techniques to account for diagnos-

risk (Heisey et al., 2014; Walsh, 2012). Given limited resources in

tic test characteristics; and (d) develop a web application to provide

wildlife management agencies, this is particularly important for dis-

easily accessible tools to implement a weighted surveillance system

ease surveillance in wildlife. For example, to improve efficiency of

for CWD. Finally, we discuss the implications of our findings in the

surveillance for chronic wasting disease (CWD), a fatal transmissible

context of CWD management, make suggestions for other agencies

spongiform encephalopathy of cervids, Walsh and Miller (2010) uti-

that are tasked with conducting wildlife disease detection surveil-

lized a weighting framework based on relative disease risk among dif-

lance, and highlight the generalizability of our approach to other

ferent demographic groups of mule deer (Odocoileus hemionus). The

disciplines.

|

Journal of Applied Ecology       3

JENNELLE et al.

2 | M ATE R I A L S A N D M E TH O DS 2.1 | Study area

peritonitis, arthritis, starvation, and nutritional deficiencies (Gilch, 2016). For assignment to the surveillance class “deer found dead,” we designated that based on external examination, carcasses must

Chronic wasting disease was discovered in south-­central

be: of adult aged deer (≥2 years old), fresh with no signs of scav-

Wisconsin, USA, in three adult males harvested in the fall of

enging or decomposition (which could mask a non-­disease related

2001. Our study area included the CWD management zone en-

source of mortality), and free of obvious physical injuries includ-

compassing ≈23,310 km2 (Figure 1a; Wisconsin Department of

ing broken bones and wounds sustained by a projectile(s) or non-­

Natural Resources, 2010). Currently, 25 of 72 Wisconsin counties

human predator(s).

have confirmed CWD in wild deer, and apparent prevalence in

Following (Heisey et al., 2014), we used a discrete proportional

adult males has increased to 50% in some local areas (Wisconsin

hazards model to estimate the relative CWD infection rate associ-

Department of Natural Resources, 2017). Despite significant ef-

ated with the surveillance class of interest based on the Wisconsin

forts to reduce deer populations following detection, the distribu-

data:

tion and magnitude of CWD has been increasing (Hefley, Hooten,

E[yij ] = 1 − exp ( − exp ([𝜇ref + xij 𝜷 + 𝜏 × ti + 𝛼j ]))

Russell, Walsh, & Powell, 2017; Heisey et al., 2010; Jennelle et al., 2014).

(1)

where yij is the binary outcome (yij = 1 indicates a positive CWD

Shenandoah National Park is situated in the Blue Ridge

sample) for deer i located in the jth spatial unit; μ ref is an inter-

Mountains near Luray, VA, USA (Figure 1b). It spans 145 km in the

cept term for the reference class against which other surveillance

north-­south direction, and encompasses approximately 80,000 ha in

classes are compared; x ij is a row vector indicating to which surveil-

nine Virginia counties in the Central Appalachian region. CWD was

lance class animal i belongs and its location in the jth spatial unit;

first diagnosed in a free-­ranging, adult doe in northwestern Virginia

β is a column vector of surveillance class log infection rate ratios; τ

in 2009, and positive cases were subsequently detected in south-

accounts for a linear time trend in overall CWD infection rate; ti is

eastern locations, approximately 16 km from SHEN. This led to con-

the year the ith deer was collected; and α j is a random effect for the

cerns about the presence of CWD in the SHEN white-­t ailed deer

jth spatial unit. We specified the reference class (μ ref ) as hunter-­

population (National Park Service, 2014).

harvested yearling males because it is a significant portion of the annual harvest and thus a demographic class for which CWD pres-

2.2 | Estimating relative weights

ence is particularly relevant (Heisey et al., 2014; Walsh & Miller, 2010). The inverse of the complimentary log–log link (inv-cloglog)

To estimate the relative weights for the surveillance classes of

function maps parameters from the complimentary log–log scale

interest, we used 90,212 CWD sample records for white-­t ailed

to the prevalence scale such that π ref = inv-cloglog(μ ref ) is preva-

deer harvested in the CWD management zone of Wisconsin from

lence of the reference class (hunter-­harvested yearling males) in

2003 to 2010 (Figure 1a). Of these, retropharyngeal lymph node

our surveillance stream, and π l = inv-cloglog(μ ref + β l) is prevalence

and/or brain stem (obex) tissues from 1,130 animals tested posi-

of the lth surveillance class. Thus, wl = exp(β l) estimates the CWD

tive for CWD using immunohistochemistry or ELISA (Keane et al.,

infection ratio of the lth surveillance class and represents the pro-

2008). These data represent the “learning dataset” of Heisey et al.

pensity of CWD to occur in the lth surveillance class relative to

(2014). We classified animals into 16 discrete surveillance classes

the reference class. When the number of disease cases is relatively

based on age, sex, and mortality source. We categorized age into

large (as in our data), wl are essentially equivalent to the infection

three groups; fawns (≈4 months–1 year), yearlings (≈1–2 year),

risk ratio (Heisey et al., 2014). It is worth noting that π ref refers to

and adults (>2 year) and assigned each harvested animal to either

apparent prevalence in the reference class. Apparent prevalence

the section (≈2.59 km2) or township (≈93.2 km2) scale using the

is derived from the force-­of-­infection and disease-­associated mor-

Public Land Survey System (US Department of the Interior, 2009).

tality rates, and is the proportion of samples in the surveillance

Mortality sources included hunter harvest, sharpshooting, vehi-

stream that test positive for CWD (Heisey, Joly, & Messier, 2006).

cle collision, clinical suspects with signs of CWD infection, and

Hereafter, for brevity, we refer to apparent prevalence as simply

deer found dead (Table 1). We defined clinical suspects as adult

prevalence.

aged deer (≥2 years old) exhibiting physical signs (excessive weight

We used WinBUGS (Lunn, Thomas, Best, & Spiegelhalter, 2000)

loss; poor coat condition; aspiration pneumonia evidenced by dif-

to solve this regression using a Laplace prior beta(1,1) on π ref, nor-

ficulty breathing or swallowing, coughing, and discharge from

mal(0, 0.0001) priors on βl and τ (parameterized with the mean and

the nasal passages), behavioural abnormalities (loss of wariness;

precision), and an intrinsic Gaussian conditional autoregressive

somnolence; excessive salivation, drinking, or urination; low head

(ICAR) structure (Osnas, Heisey, Rolley, & Samuel, 2009; Wakefield,

carriage; drooped ears), and/or locomotor conditions (ataxia; head

Best, & Waller, 2001; Waller, Carlin, Xia, & Gelfand, 1997) for the

tremors) consistent with CWD infection (Gilch, 2016). These signs,

spatial random effect αj. We used a uniform (0, 100) distribution for

while nonspecific, increasingly manifest with progression of the

the hyperprior standard deviation with the ICAR structure. We ex-

disease and could potentially mimic symptoms of other conditions

amined prior sensitivity using both smaller and larger variances for

like brain abscesses, traumatic injuries, meningitis, encephalitis,

the priors, and in all cases, the resulting parameter estimates were

|

Journal of Applied Ecology 4      

JENNELLE et al.

(a)

(b)

F I G U R E   1   (a) Chronic wasting disease management zone in Wisconsin, USA, used to collect the training dataset to estimate relative chronic wasting disease detection weights for different surveillance classes of white-­t ailed deer. (b) Shenandoah National Park study area in Virginia, USA, used for the application of the weighted surveillance approach to chronic wasting disease detection in white-­t ailed deer.

similar. We initialized three Markov chains with differing starting

(Brooks & Gelman, 1998), and inspected trace and autocorrelation

values within the acceptable parameter space, and used a burn-­in

plots of the parameters. All parameter estimates were derived using

period of 70,000 iterations. To assess convergence of posterior

the final 30,000 iterations from the posterior distributions of each of

distributions, we calculated the Brooks–Gelman–Rubin statistic

the three chains (Supporting Information Appendix S1).

|

Journal of Applied Ecology       5

JENNELLE et al.

assuming a Laplace prior for π ref (Heisey et al., 2014), and using

2.3 | Surveillance weights

the posterior mean and precision for βl estimated above, as the mean

Prior to estimating the potential contribution of each surveillance

and precision for an informative normal prior on βl. Finally, from this

class (Table 1) to detecting the presence of CWD (i.e., “real weights”

posterior distribution we estimated pdet. We repeated this proce-

of Heisey et al. 2014), a detection surveillance goal must be deter-

dure for various sample sizes (Nl) of the lth surveillance class until we

mined. It is composed of two quantities; a disease detection thresh-

determine the smallest value of Nl, which achieves pdet ≈ 0.95. This

old or reference class population prevalence (π ref ) below which the

value of Nl is then used in the following equation to calculate the

entire population is effectively disease free, and a detection thresh-

weight for the lth surveillance class:

old probability (pdet; termed β in Heisey et al. 2014). Generally, π ref is set to 0.01 and pdet is set to 0.95, and the surveillance goal is to col-

(3)

Rl = 298∕Nl

lect enough samples to ensure there is a 95% probability the disease prevalence is below 1% in the reference population. This goal can be

These estimates describe the approximate number of reference

restated probabilistically as Pr(π ref ≤ 0.01) ≥ pdet, and requires that

class samples that an individual from the lth surveillance class would

CWD is not detected in 298 randomly selected individuals from the

be worth. For example, if Rl = 10 an individual from the lth surveil-

reference class (Walsh & Miller, 2010). The methodology is the same

lance class provides the same amount of disease detection informa-

regardless of quantities chosen for π ref or pdet (e.g., 2.5% for SHEN

tion as approximately 10 individuals from the reference class (Heisey

described below).

et al., 2014; Walsh & Miller, 2010). While these weights (Rl) are not

We estimated the weights for each surveillance class using the

purely additive (Heisey et al., 2014), they can be used to adequately

target bound-­matching approach of (Heisey et al., 2014). To esti-

approximate the number of reference samples needed to reach sur-

mate the weight (Rl) for a nonreference surveillance class l (recall

veillance sample goals.

the weight for the reference class ≡ 1), we assumed all surveillance samples arise from the lth surveillance class, and CWD is undetected. Next, we specified a starting sample size (Nl) for the lth

2.4 | SHEN case study

surveillance class (as explained below, this is an iterative process).

Given the risk of CWD introduction into SHEN, we conducted a CWD

Then, we estimated the posterior distribution of π ref using the fol-

surveillance programme of the deer population within the northern

lowing likelihood:

half of the park. We used yearling male deer as the reference class and weights estimated from the Wisconsin surveillance data to de(2)

L(𝜇ref ) = 𝜋l0 (1 − 𝜋l )Nl = (1 − inv cloglog(𝜇ref + 𝛽l ))Nl

termine the number of adult (≥2 years) male and adult female deer needed to detect the presence of CWD in SHEN with a 95% probabil-

TA B L E   1   Classification of surveillance classes designated for harvested white-­t ailed deer within the CWD management zone of Wisconsin from 2003 to 2010. These classes represent all CWD risk groups considered in this analysis. N and I are the total and infected numbers of deer in each class, respectively Mortality source Hunter-­harvested fawn female

4021 (5) 13,694 (81)

Hunter-­harvested adult female

27,091 (267) 3,690 (1)

Hunter-­harvested yearling male

17,442 (115)

Hunter-­harvested adult male

20,560 (510)

Sharpshot fawn female

299 (4)

Sharpshot yearling female

256 (2)

Sharpshot adult female

1,053 (27)

Sharpshot fawn male

358 (7)

Sharpshot yearling male

323 (4)

Sharpshot adult male

351 (21)

Vehicle collision (direct or indirect) Found dead

sensitivity and specificity of the diagnostic test. Given the 2.5% detection threshold, 118 yearling males were required; we refer to this quantity as reference unit equivalents (E). We chose a 2.5%, rather than 1% prevalence target, based on limited available resources. We calculated the number of samples needed from surveillance

N (I)

Hunter-­harvested yearling female Hunter-­harvested fawn male

ity if prevalence was ≤0.025 in yearling males, and assuming a 100%

445 (2)

classes based on various combinations of adult males and females using the following equation: nfemale =

E − (nmale × Rmale ) Rfemale

,

(4)

where nfemale is the number of adult females, nmale is the number of adult males, and Rmale or Rfemale represent the weights from the Wisconsin CWD data for adult males and females, respectively. We used these combinations as potential quotas for our CWD-­detection surveillance program. However, samples from other risk classes (e.g., vehicle-­killed deer) also entered the surveillance stream. Therefore, we updated our sample quota in real time based on the following inequality:

E−

I ∑

(ni × Ri ) ⩽ 0.

(5)

i=1

91 (11)

Clinical suspect; hunter reported

368 (33)

We captured deer using a combination of butorphanol (~0.5 mg/kg),

Clinical suspect; community reported

170 (40)

azaperone (~0.2 mg/kg), and medetomidine (~0.25 mg/kg) (BAM,

|

Journal of Applied Ecology 6      

JENNELLE et al.

TA B L E   2   Chronic wasting disease (CWD) log infection ratios (β), infection ratios (w), surveillance weights (R), approximate number of samples (n) needed to achieve Pr(π ref ≤ 0.01) ≥ 0.95, and precision (SD) from using white-­t ailed deer harvest data from 2003 to 2010 in the CWD management zone of Wisconsin. Hunter-­harvested yearling males are the reference class Mortality source

β (SD)

w (SD)

R

n

Clinical suspect; community reported

3.549 (0.208)

35.530 (7.426)

33.333

9

Clinical suspect; hunter reported

2.270 (0.209)

9.888 (2.062)

9.091

33

Found dead

2.204 (0.337)

9.575 (3.198)

7.317

41

Sharpshot adult male

1.294 (0.246)

3.757 (0.9193)

3.297

91

Hunter-­harvested adult male

1.169 (0.105)

3.237 (0.3433)

3.226

93

Sharpshot adult female

0.527 (0.224)

1.736 (0.3895)

1.563

192

1.328 (0.1508)

1.304

230

1.000

298

Hunter-­harvested adult female Hunter-­harvested yearling male

0.277 (0.113) —

—

Sharpshot fawn male

−0.144 (0.410)

0.939 (0.378)

0.625

480

Vehicle collision

−0.381 (0.806)

0.898 (0.641)

0.216

1,391

Hunter-­harvested yearling female

−0.142 (0.147)

0.877 (0.130)

0.850

353

Sharpshot yearling male

−0.279 (0.548)

0.868 (0.449)

0.432

695

Sharpshot fawn female

−0.534 (0.539)

0.669 (0.337)

0.347

865

Sharpshot yearling female

−0.957 (0.808)

0.506 (0.367)

0.121

2,475

Hunter-­harvested fawn female

−2.026 (0.485)

0.147 (0.068)

0.084

3,570

Hunter-­harvested fawn male

−4.016 (1.276)

0.032 (0.032)

0.001

250,600

1.224 (0.0202)

—

—

τ – year effect

0.202 (0.017)

Wildlife Pharmaceuticals, Windsor, CO, USA) as a chemical immo-

Lastly, based on our surveillance results, we estimated the ex-

bilizing agent. We used hand-­held dart guns (Dan-­Inject, Austin, TX,

pected prevalence limits of CWD in SHEN for three demographic

USA) with barbed, radiotransmitter darts (Pneudart, Williamsport, PA,

classes (adult males, adult females, and yearling males) of interest

USA) for drug delivery. We focused sampling in less mountainous areas

to park personnel. We used the following likelihood, in combina-

of SHEN because these regions hosted the largest deer populations,

tion with appropriate prior distributions, to estimate the posterior

which we assumed would correlate with a higher likelihood of detect-

distribution of π l (class-­specific prevalence), which allows for im-

ing CWD if it was present. An experienced National Park Service (NPS)

perfect diagnostic tests in contrast to previous methods (Heisey

biologist estimated age of captured deer using tooth eruption and wear

et al., 2014):

patterns. After collecting rectoanal mucosa-­associated lymphoid tissue (RMALT) biopsy samples from each captured individual (Wolfe et al., 2007), anaesthesia was reversed using 40–50 mg of atipamezole, 300– 400 mg tolazoline, and 100 mg naltrexone (Wildlife Pharmaceuticals) delivered intramuscularly. Capture and handling was approved by the NPS IACUC (ID:NER_SHEN_Walsh_WtD_2012). We included vehicle-­ killed deer within and adjacent to the park in our sampling stream and retropharyngeal lymph nodes were collected from these animals and tested for CWD. We fixed tissues in 10% neutral buffered formalin and shipped them to Colorado State University (Fort Collins, CO, USA) for a CWD assay. Immunohistochemically stained tissue sections were evaluated to determine the presence of the protease-­resistant prion protein indicative of infection with CWD (Peters, Miller, Jenny, Peterson, & Carmichael, 2000; Wolfe et al., 2007). We initially assumed our diagnostic tests of RMALT were perfect (Heisey et al., 2014); however, recent work shows that test sensitivity (Se) is lower when using RMALT for CWD (Thomsen et al., 2012). Therefore, post hoc, we adjusted E to account for Se using the following equation:

L(𝜋l , Sel , Spl |Cl , nl ) = [(𝜋l × Sel ) + (1 − 𝜋l ) × (1 − Spl )]Cl [𝜋l × (1 − Sel ) + (1 − 𝜋l ) × (Spl )]nl −Cl

(7)

where C l is the number of CWD positive tests out of n l samples from the lth surveillance class, π l = inv-cloglog(π ref + β l), Se l is test sensitivity, and Sp l is test specificity (assumed 1 hereafter). These latter two variables can also vary between surveillance classes (l) if different diagnostic tests are used. As before, we used a Laplace prior for π ref and the posterior mean and precision for β l estimated from the Wisconsin data, as an informative prior. This allowed us to account for different test sensitivity for antemortem RMALT samples and postmortem retropharyngeal lymph nodes. We specified an informative beta (92.5, 44.5) prior for animals sampled via RMALT based on an estimated sensitivity of 68% with a 95% CI of 49%–82% (Thomsen et al., 2012). For vehicle-­k illed deer, we specified an informative beta (73.5, 1.5) prior for test sensitivity based on (Miller & Williams, 2002). We present the 95% upper credible bounds from our posterior distributions for the potential prevalence in the park because these provide the

Eadj =

ln (1 − pdet ) ln (1 − 𝜋ref × Se)

(6)

most useful summary of potential CWD risk (i.e., 95% probability the latent prevalence rates are below the upper credible bound).

|

Journal of Applied Ecology       7

JENNELLE et al.

The WinBUGS code used for this final analysis is provided in

4 | D I S CU S S I O N

Supporting Information Appendix S2. Additionally, we have created a web application (https://popr.cfc.umt.edu/CWD/) that

Extending upon techniques in Heisey et al. (2014), we quantified

conducts these analyses as well as assists in designing a weighted

relative CWD risk of various surveillance classes of white-­t ailed deer

surveillance system and allows for real-­t ime tracking of sampling

in Wisconsin. Our findings are similar to those reported for mule

quotas (Equation 6).

deer in Colorado, with the highest infection risk in deer exhibiting clinical signs of CWD, and increasing risk for older deer, especially

3 | R E S U LT S 3.1 | Wisconsin data and surveillance weights

males (Heisey et al., 2014; Walsh & Miller, 2010). However, we used a greater number of surveillance classes (l = 16) in this study (e.g., found dead deer) to broaden the sources of deer mortality that CWD surveillance programmes might encounter. To demonstrate

Based on graphical analysis and insignificant interactions between

the range of applicability of these surveillance classes for reaching

year and class-­specific infection rate, we deemed our proportional

surveillance goals, we provide example sample sets that achieve

hazards model appropriate for use. In Wisconsin, clinical suspect

Pr(π ref ≤0.01) ≥0.95 (Supporting Information Table S1). Additionally,

deer and deer found dead exhibited the highest CWD prevalence

we estimated surveillance weights for each class (relative to a refer-

relative to harvested yearling males, while car-­killed deer and fawns

ence), which accounted for uncertainty in the estimates of relative

exhibited the lowest infection ratios (Table 2). Adult animals and

CWD risk. We demonstrated how weights from Wisconsin could

males had higher CWD prevalence than younger individuals and fe-

be used to inform CWD detection surveillance in SHEN including

males, respectively. Preliminary modelling indicated no interaction

sampling effort, sample sizes, and resources. These weights can also

between time and mortality source effects, and we did not detect

be used for planning future surveillance activities in other regions

broad scale spatial differences within the management zone. We de-

where early detection of CWD is important to guide management

tected a statistically significant time trend in CWD infection in the

strategies. The practical significance of using these weights is the

management zone during 2003–2010 (Table 2) with a 95% probabil-

ability to target surveillance classes that are most likely to be CWD-­

ity of increasing annual rates between 18% and 26%. The number of

infected, which can substantially reduce the number of samples re-

samples needed to achieve a surveillance goal of Pr(π ref ≤ 0.01) ≥0.95

quired to reach a given surveillance goal.

ranged from nine clinical suspect deer to 250,600 hunter-­harvested male fawns (Table 2).

Although CWD was not detected in SHEN, we estimated the posterior distribution of the expected CWD prevalence limits for three demographic groups of interest, given our sampling effort.

3.2 | SHEN case study

If CWD is present in SHEN, these estimates provide a detectable range of CWD prevalence from our surveillance sample. An alter-

We used the weights for hunter-­harvested adult males (3.226) and

native explanation of the results arises due to the equivalence of

hunter-­harvested adult females (1.304) to determine the necessary

hypothesis testing and credible intervals (Heisey et al., 2014), for ex-

sample sizes for our surveillance efforts in SHEN, assuming relative

ample, 95% upper credible bounds for prevalence can be interpreted

CWD prevalence was similar to that in Wisconsin. Assuming perfect

as follows: if CWD is present in SHEN and given our sampling ef-

diagnostic test sensitivity, our surveillance goal of Pr(πref ≤ 0.025)

fort, we are 95% confident that we would have detected at least one

≥0.95 required 118 samples from hunter-­harvested yearling males

CWD-­infected animal if the prevalence in adult males, adult females,

or 37–91 adult male and/or adult female deer with fewer animals required as the number of adult males in the surveillance stream increased (Table 3). We captured and tested 70 deer (21 adult males and 49 adult females) in 2012–2013 from the northern half of SHEN. At capture, male ages ranged from 2.5 to 8.5 years with 16 adult males

TA B L E   3   The approximate number of samples required from the adult male and adult female demographic groups of white-­t ailed deer in Shenandoah National Park, Luray, VA, USA to detect chronic wasting disease at a prevalence of 0.025 in yearling males with a 95% probability (assuming perfect test diagnostics)

≤3.5 years, and female ages ranged from 2.5 to 8.5 years with 30

Adult female samples

Adult male samples

Total samples

females ≤3.5 years. We sampled 10 vehicle-­killed deer during that

0

37

37

time period. Eleven deer (three males and eight females) lacked

5

35

40

sufficient rectal tissue (≥3 follicles required; Keane et al., 2009 rec-

17

30

47

ommend ≥6) to allow for CWD testing and could not be included in subsequent analyses. Prion infection was not detected in any deer (i.e., Ci = 0 for all classes), although the 95% upper credible bounds from the posterior distributions for CWD prevalence, if present in SHEN, were 0.112 for adult males, 0.048 for adult females, 0.041 for vehicle-­killed deer, and 0.036 for yearling males (i.e., baseline class).

29

25

54

42

20

62

54

15

69

67

10

77

79

5

84

91

0

91

|

Journal of Applied Ecology 8      

JENNELLE et al.

vehicle-­killed deer, and yearling males is at least 11.2%, 4.8%, 4.1%

detecting CWD ≥ 90%). Therefore, unless systems are markedly dif-

or 3.6%, respectively. Our sampling effort was not sufficient to

ferent, using surveillance weights from one region to conduct CWD

achieve our surveillance goal of Pr(π ref ≤0.025) ≥0.95, and under-

surveillance in a separate region may not significantly reduce the

scores the challenge of establishing an effective surveillance stream.

probability of detecting a CWD positive animal below the nominal

Our sample shortfall was partly due to 11 unusable samples having

level. However, we advise caution applying our weights for clinical

insufficient follicles in RMALT biopsy tissues. Additionally, because

suspects and deer found dead when it is known or suspected that

of imperfect diagnostic test sensitivity for the RMALT procedure

hard winters or other disease conditions mimicking late stage CWD

(average = 0.68) that we used to test live deer for CWD in SHEN, the

infection (such as haemorrhagic diseases) may be causing localized

required reference unit equivalents (E) to achieve our surveillance

or regional deer mortality. In either of these cases, it is preferable to

goal increased from 118 (perfect test sensitivity) to 175 harvested

apply the appropriate hunter-­harvest weights we estimated.

yearling males. This highlights a useful aspect of the weighted sur-

An additional consideration is that our surveillance work at

veillance approach; sampling calculations can be updated while sur-

SHEN was conducted over approximately a year. Therefore, we as-

veillance activities are ongoing. Thus, effort and surveillance class

sumed the prevalence of CWD, if present in SHEN, did not change

sources can be accounted for in real-­time to determine additional

appreciably over the time period during which we collected our

samples and resources needed to achieve surveillance goals. This is

samples (