Abstract

Understanding the distribution of pathogens across landscapes and their prevalence within host populations is a common aim of wildlife managers. Despite the need for unbiased estimates of pathogen occurrence and prevalence for planning effective management interventions, many researchers fail to account for imperfect pathogen detection. Instead raw data are often reported, which may lead to ineffective, or even detrimental, management actions. We illustrate the utility of occupancy models for generating unbiased estimates of disease parameters by 1) providing a written tutorial describing how to fit these models in Program PRESENCE and 2) presenting a case study with the pathogen ranavirus. We analyzed ranavirus detection data from a wildlife refuge (Maryland, US) using occupancy modeling, which yields unbiased estimates of pathogen occurrence and prevalence. We found ranavirus prevalence was underestimated by up to 30% if imperfect pathogen detection was ignored. The unbiased estimate of ranavirus prevalence in larval wood frog (Lithobates sylvaticus; 0.73) populations was higher than in larval spotted salamander (Ambystoma maculatum; 0.56) populations. In addition, the odds of detecting ranavirus in tail samples were 6.7 times higher than detecting ranavirus in liver samples. Therefore, tail samples presented a nonlethal sampling method for ranavirus that may be able to detect early (nonsystemic) infections.

INTRODUCTION

Emerging infectious diseases (EIDs) in wildlife populations are an increasingly common threat to biodiversity (Daszak et al. 2000; Fisher et al. 2012). Developing appropriate management responses to EIDs is an important goal for many natural resource managers and researchers (Grant et al. 2017; Russell et al. 2017; Gerber et al. 2018). Assessing disease risks and implementing management interventions to improve a population's resilience to disease requires reliable understanding of a pathogen's distribution at multiple spatial and temporal scales (McClintock et al. 2010).

The coarsest scale of interest to resource managers is often where a pathogen occurs on the landscape and which factors shape that distribution through space and time (Adams et al. 2010; Lorch et al. 2013). The focus is on estimating the proportion of sites (i.e., discrete habitat patches) occupied by infected host populations, or the probability that a randomly selected site with certain characteristics is occupied by an infected host population. It is possible to estimate changes in the distribution of infected host populations through time, using multiyear studies, giving rise to estimates of the pathogen's local transmission and extinction rates. Investigating shifts in a pathogen's spatial or temporal distribution in hosts is useful for understanding the pathogen's ecology and origins (Rachowicz et al. 2005), assessing host species extinction risk (Rödder et al. 2009), and predicting host-pathogen dynamics under future climate and management scenarios (Rogers and Randolph 2000; Gerber et al. 2018).

Researchers and managers may also seek to understand within-population disease dynamics. In these cases, the focus is on variation in the proportion of individuals infected in a population (i.e., prevalence), and in the characteristics (e.g., stage or age, sex, size) that may influence infection probability. Comparing prevalence estimates for a single host population through time can provide insights about host impacts, the transmission process, and the spread of disease (Caley and Ramsey 2001). Alternatively, comparing estimates of prevalence across populations and species yields information about habitat-mediated disease dynamics, differential susceptibility (Scheele et al. 2017), and dynamics of the disease and its transmission in multi-host systems (Keesing et al. 2006). Monitoring changes in prevalence can help identify enzootic from epizootic infections, which often call for different management actions (Langwig et al. 2015).

Variables such as site occurrence and prevalence cannot be observed without error; failure to detect a pathogen can occur at several stages in the observation process (Fig. 1; McClintock et al. 2010). At the landscape and site levels, a pathogen might go undetected if an occupied habitat or infected individuals are not sampled, a concern when prevalence or population size are low, or when prevalence is low and population size is large (Green et al. 2009; MacKenzie et al. 2018). The problem is amplified if infected animals are more difficult to detect than uninfected individuals (Jennelle et al. 2007; Valenzuela-Sánchez et al. 2017), or more likely to permanently emigrate (Schmidt 2010). Pathogen occurrence may go undetected at the individual level either because the sample collected (e.g., a skin swab, tail clip) did not contain the pathogen or because the diagnostic test used was imperfect (i.e., low sensitivity; Thompson 2007). For instance, molecular assays can fail to amplify pathogens due to low quantities of DNA or the presence of assay inhibitors (McKee et al. 2015; Goldberg et al. 2016). All detection methods are imperfect to some extent (Lachish et al. 2012; Miller et al. 2012), and this must be considered when inferences on occurrence and prevalence are to be used for management strategies or for increased ecologic understanding.

Figure 1

Visual depiction of the multiple scales of interest in a ranavirus case study at the Patuxent Research Refuge (Laurel, Maryland, USA). Wetlands and tadpoles are shaded black or white to represent the true occupancy state of ranavirus, where black indicates occurrence and white indicates absence. Questions of interest are related to pathogen occurrence (within wetlands, ψsite, or individuals, ψind or θ, and/or pathogen detection (ρsite, ρind, or ρmult) based on detection (+) and nondetection (_) data from one liver (L) and two tail (T1, T2) samples per individual. Detection/nondetection data from the laboratory are converted to 0s and 1s for modeling purposes. False positives do not occur in our system, but ranavirus may be detected imperfectly as depicted. We present a site occurrence analysis to estimate site occurrence and detection probability, a prevalence analysis to estimate prevalence and detection probability, and a multiscale analysis to estimate all quantities simultaneously. Data consist of 0s and 1s indicating whether or not ranavirus was identified in each sample type.

Figure 1

Visual depiction of the multiple scales of interest in a ranavirus case study at the Patuxent Research Refuge (Laurel, Maryland, USA). Wetlands and tadpoles are shaded black or white to represent the true occupancy state of ranavirus, where black indicates occurrence and white indicates absence. Questions of interest are related to pathogen occurrence (within wetlands, ψsite, or individuals, ψind or θ, and/or pathogen detection (ρsite, ρind, or ρmult) based on detection (+) and nondetection (_) data from one liver (L) and two tail (T1, T2) samples per individual. Detection/nondetection data from the laboratory are converted to 0s and 1s for modeling purposes. False positives do not occur in our system, but ranavirus may be detected imperfectly as depicted. We present a site occurrence analysis to estimate site occurrence and detection probability, a prevalence analysis to estimate prevalence and detection probability, and a multiscale analysis to estimate all quantities simultaneously. Data consist of 0s and 1s indicating whether or not ranavirus was identified in each sample type.

Underestimating pathogen distributions could result in management actions such as reintroductions occurring in suboptimal locations (Muths et al. 2014) and makes predicting pathogen spread difficult, leaving managers to deal with more expensive reactive control measures rather than proactive prevention measures (Johnson et al. 2015). Underestimates of prevalence, stemming from a failure to diagnose individuals as infected, might delay the initiation of management actions like culling, visitor-use restrictions, or chemical treatments (Wasserberg et al. 2009; Woodhams et al. 2011; Langwig et al. 2015). Despite the need for unbiased estimates of host-pathogen disease dynamics, naïve detection data (that do not account for imperfect and variable detection probabilities) are often the only metric reported in pathogen distribution studies.

Occupancy models allow for estimation of pathogen occurrence while explicitly accounting for imperfect and variable detection probabilities (MacKenzie et al. 2002; Tyre et al. 2003). The models account for false negative detection errors by relying on multiple surveys over space or time during a window where the true pathogen occurrence status of the study unit (a site or an individual) is closed to changes (MacKenzie et al. 2017).

Our goal was to illustrate an analytical framework for estimation of unbiased pathogen parameters of a wildlife disease using a standard sampling design and a case study on amphibian ranavirus in the northeastern US. We used three occupancy analyses of pathogen detection data to demonstrate the estimation of site-level occurrence of infected populations, individual-level occurrence (prevalence) of pathogens, or both quantities simultaneously. We discuss how our results inform ranavirus monitoring and surveillance programs. We focused on a ranavirus case study but these methods are readily transferrable to other pathogen systems, and we provide a tutorial (see Supplementary Materials) to help readers fit these models to data from other systems.

Ranaviruses (Family Iridoviridae) are generalist pathogens that have been linked to amphibian die-offs and population declines (Daszak et al. 2003; Duffus et al. 2015). Larval amphibians are especially susceptible to ranavirus infections (Hoverman et al. 2011), although adults can be negatively affected as well (Price et al. 2014; Sutton et al. 2014). Ranaviruses have infection affinity for vascular epithelial cells, and cell death results in widespread organ necrosis, failure of blood vessels, and internal hemorrhaging (Chinchar 2002; Miller et al. 2015). Ranavirus-induced declines can be severe; mortality approaches 100% at some locations (Green et al. 2002; Price et al. 2014; Wheelwright et al. 2014). Host death can occur within only a few days of infection and die-offs may occur rapidly (Hoverman et al. 2011; Wheelwright et al. 2014). Because mortality occurs around the same time as metamorphosis and amphibians decompose or are scavenged rapidly (Brunner et al. 2015), it can be difficult to determine whether nondetection of larval amphibians on a late-season survey is due to a ranavirus die-off or successful metamorphosis.

Although a gold standard tissue type has not been identified by the World Organization for Animal Health for ranavirus testing in amphibians (Miller et al. 2015), liver, spleen, and kidney tissues are most commonly used. Ranavirus detection in the liver generally indicates systemic infection (Miller et al. 2015), but tail clips are a common nonlethal sample assumed to have a lower detection probability than liver samples (Gray et al. 2012; Sutton et al. 2015).

MATERIALS AND METHODS

Ranavirus survey field methods

Our study was conducted at the US Fish and Wildlife Service's 5,200-ha Patuxent Research Refuge (PRR) in central Maryland, US (39°1′N, 76°47′W). The refuge is covered in bottomland hardwood and mixed forest, and is estimated to contain around 2,250 vernal pool breeding sites (Van Meter et al. 2008). We sampled 22 randomly selected wetlands for larval amphibians about 60 d after egg mass deposition when we expected to find late-stage tadpoles and larvae. Each wetland was sampled on a single day from 24 May 2011 to 26 May 2011. We used dip nets to capture wood frog (Lithobates sylvaticus) tadpoles and spotted salamander (Ambystoma maculatum) larvae. Each animal was isolated and handled with a new pair of sterile nitrile gloves to reduce the chance of cross-contamination of ranavirus. The pH of each wetland was also measured using a PC Testr 35 Multiparameter meter (Oakton Instruments, Vernon Hills, Illinois, USA).

Captured animals were immediately humanely euthanized via transdermal exposure to benzocaine hydrochloride, and placed into individual sterile plastic bags containing 95% ethyl alcohol (Green et al. 2009). Animals were sent to the University of Tennessee Center for Wildlife Health (Knoxville, Tennessee, USA) where they were necropsied; three samples were collected from each specimen: one from the liver and two from the tail.

To detect ranavirus, we followed the protocol of Sutton et al. (2014) targeting a 70-base pair region of the virus's major capsid protein and running the quantitative PCR for 40 cycles. This method detects the presence of ranavirus DNA within the samples collected, and does not detect the presence of disease. Only samples with a threshold cycle <33 were considered positive based on a standard curve of known ranavirus quantities (Miller et al. 2015). Each quantitative PCR assay included two positive controls (cultured virus and genomic DNA from a known positive tadpole) and two negative controls (DNA-free water and genomic DNA from a known negative tadpole).

Statistical methods and terminology

Researchers may be interested in the probability of pathogen occurrence in a site (ψsite), in an individual (ψind or θ), or both. These parameters are binomial random variables. We first illustrate how these quantities can be estimated in separate analyses using standard occupancy models (Mac-Kenzie et al. 2002). Then, we demonstrate how the data can be analyzed in a multiscale occupancy model (Nichols et al. 2008) to estimate both occurrence and prevalence simultaneously. These methods account for false negative detection errors, and assume that false positive detections do not occur.

Site occurrence estimation

The naïve estimate of pathogen site occurrence (ψnaive) is the proportion of wetlands where a pathogen was detected. This estimate will be biased low if detection (p) is imperfect. Alternatively, one can use replicate pathogen surveys to account for imperfect detection (ψsite). Replicate pathogen surveys may correspond to multiple environmental (e.g., environmental DNA [eDNA]; Mosher et al. 2017), host-based (e.g., skin swab; DiRenzo et al. 2018), or sub–host-based (e.g., tissue; Elmore et al. 2014) samples and must be collected over a time period when the site is assumed to be closed to changes in pathogen occurrence (Table 1).

Table 1

Example of data formatting for estimating the proportion of sites in which a pathogen occurs (site occurrence) using an occupancy framework. Each row corresponds to a sampled site and the values within a row correspond to the results from replicate samples (S1, S2, and S3) collected from each individual animal. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Occupancy models can accommodate missing observations. For example, no individuals were captured at site D, so all observations for that site are missing and coded with a dash (—). Species identity and sample type can also be included as covariates, but are not shown here for simplicity.

Example of data formatting for estimating the proportion of sites in which a pathogen occurs (site occurrence) using an occupancy framework. Each row corresponds to a sampled site and the values within a row correspond to the results from replicate samples (S1, S2, and S3) collected from each individual animal. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Occupancy models can accommodate missing observations. For example, no individuals were captured at site D, so all observations for that site are missing and coded with a dash (—). Species identity and sample type can also be included as covariates, but are not shown here for simplicity.
Example of data formatting for estimating the proportion of sites in which a pathogen occurs (site occurrence) using an occupancy framework. Each row corresponds to a sampled site and the values within a row correspond to the results from replicate samples (S1, S2, and S3) collected from each individual animal. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Occupancy models can accommodate missing observations. For example, no individuals were captured at site D, so all observations for that site are missing and coded with a dash (—). Species identity and sample type can also be included as covariates, but are not shown here for simplicity.

The probability of obtaining detection history 00110 over five surveys (where 0s and 1s denote pathogen nondetection and detection, respectively) is as follows:

formula

This history indicates that the pathogen was present (ψsite), was undetected on surveys 1 and 2 (1-psite)2, was detected on surveys 3 and 4 (psite2), and was again undetected on survey 5 (1-psite). Sites with all zero detection histories (i.e., 00000) are ambiguous because observations could arise in two ways; either because the pathogen is truly absent (1-ψsite) or because the pathogen was present but went undetected on all visits (ψsite[1-psite]5). Information from sites where both detections (value=1) and nondetections (value=0) occur help to resolve these ambiguities (MacKenzie et al. 2002).

Prevalence estimation

When multiple individuals are sampled for a pathogen at a single site, the resulting data can be used to investigate the prevalence (ψind) of the pathogen in that population using the same occupancy model described above (Table 2). In this case, prevalence is the probability that a pathogen occurs in an individual rather than in a site.

Table 2

Example of data formatting for estimating the infection prevalence of a pathogen at a single site of interest using occupancy models. Each row corresponds to a sampled individual and the values within a row correspond to the results from three samples collected. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Species identity and tissue sample type can also be included as covariates, but are not shown here for simplicity.

Example of data formatting for estimating the infection prevalence of a pathogen at a single site of interest using occupancy models. Each row corresponds to a sampled individual and the values within a row correspond to the results from three samples collected. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Species identity and tissue sample type can also be included as covariates, but are not shown here for simplicity.
Example of data formatting for estimating the infection prevalence of a pathogen at a single site of interest using occupancy models. Each row corresponds to a sampled individual and the values within a row correspond to the results from three samples collected. Samples where the pathogen was detected are coded as a “1″ while nondetections are coded as a “0.” Species identity and tissue sample type can also be included as covariates, but are not shown here for simplicity.

Combined analysis

Pathogen occurrence (at sites), prevalence (within populations), and detection (within samples) compose a hierarchical research question (Fig. 1). A multiscale extension of the standard occupancy model (Nichols et al. 2008) allows for estimation of all three parameters simultaneously and avoids nonindependence issues that may be present in other formulations. Traditional singleseason occupancy models assume closure across all observations at a site during the survey period (in this case, that 100% of amphibians at a site are infected or uninfected). The multiscale model alleviates this assumption by allowing some animals to be infected while others are not. This model was designed to account for variation in occurrence across space or time, but the concept is also readily applicable to sites and individuals in a pathogen occurrence framework (Schmidt et al. 2013; Elmore et al. 2016; Mosher et al. 2017). Conditional on pathogen occurrence at the site (ψsite), some proportion of individuals may also be infected (θ; prevalence). The probability of detecting the pathogen in a single sample (pmult) is conditional on pathogen presence in both the site and individual. Using the multiscale model, the probability of obtaining a detection history of 010000 for a site with two sampled individuals and three tissue samples per individual is as follows:

formula

This history indicates that the pathogen was present at the site (ψsite) and in the first sampled individual (θ), despite the fact that it was only detected on one tissue sample. The second individual's history (000) is ambiguous, and could have arisen because the individual was uninfected (1-θ), or because the individual was infected (θ) but the laboratory methods failed to detect the pathogen on all three trials ([1-pmult]3).

Detection probability

Each analysis (site occurrence, prevalence, multiscale) includes an estimation of detection probability, p. The interpretation of these detection probabilities varies depending on the study design and analysis, which we note by subscripting p based on the analysis being discussed. In the site occurrence analysis, the detection probability estimated (psite) is in fact a product of the true prevalence and detection probability, also called the positive predictive value (Nichols et al. 2008), rather than the diagnostic sensitivity of the assay. In the prevalence analysis, the replicates within a sample unit come from only one individual, so the detection probability (pind) is no longer confounded with prevalence, but in our example, inference is restricted to the one site examined. Finally, in the multiscale occupancy model, prevalence and detection are no longer confounded, and pmult is the diagnostic sensitivity of the assay.

Software and multi-model inference

We used all of the ranavirus data (22 wetlands) for the site occurrence and multiscale analyses, and used data from one wetland (with 24 sampled individuals) for the prevalence analysis. We generated a set of biologic hypotheses for each analysis (Table 3), and evaluated the strength of those hypotheses by fitting competing models. Models were fit using single-season standard and multiscale occupancy models in Program PRESENCE (Hines 2006). We evaluated support for competing hypotheses using Akaike's information criterion adjusted for small sample sizes (AICc) and associated model weights (Burnham and Anderson 2003).

Table 3

Biologic hypotheses related to ranavirus occurrence (ψsite), prevalence (ψind, θ), and detection probability (p) in a study of wood frog (Lithobates sylvaticus) tadpoles and spotted salamander (Ambystoma maculatum) larvae at the Patuxent Research Refuge (Laurel, Maryland, USA). The covariates listed for ψsite were used in the site occurrence and multiscale models, the covariates listed for ψind, θ were used in the prevalence and multiscale models, and the covariates listed for p were used in the site occurrence, prevalence, and multiscale occupancy models.

Biologic hypotheses related to ranavirus occurrence (ψsite), prevalence (ψind, θ), and detection probability (p) in a study of wood frog (Lithobates sylvaticus) tadpoles and spotted salamander (Ambystoma maculatum) larvae at the Patuxent Research Refuge (Laurel, Maryland, USA). The covariates listed for ψsite were used in the site occurrence and multiscale models, the covariates listed for ψind, θ were used in the prevalence and multiscale models, and the covariates listed for p were used in the site occurrence, prevalence, and multiscale occupancy models.
Biologic hypotheses related to ranavirus occurrence (ψsite), prevalence (ψind, θ), and detection probability (p) in a study of wood frog (Lithobates sylvaticus) tadpoles and spotted salamander (Ambystoma maculatum) larvae at the Patuxent Research Refuge (Laurel, Maryland, USA). The covariates listed for ψsite were used in the site occurrence and multiscale models, the covariates listed for ψind, θ were used in the prevalence and multiscale models, and the covariates listed for p were used in the site occurrence, prevalence, and multiscale occupancy models.

RESULTS

We collected up to 12 individuals of each species per site (Fig. 1). This sample size was optimal (based on Monte Carlo simulations) given our joint objectives of maximizing the precision of both prevalence and site occurrence estimates simultaneously. We collected a total of 321 larval amphibians (117 spotted salamander larvae and 204 wood frog tadpoles). Two of the collected animals were recently deceased, while all others were alive. Without accounting for imperfect detection, the naïve estimate of ranavirus prevalence was 42% for each species (85 of 204 wood frogs and 49 of 117 spotted salamanders). Naïve ranavirus site occurrence was 50% (detections at 11 of 22 sites).

Site occurrence

The best-supported model (Table 4; weight=0.81) suggested that ranavirus occurrence was constant across the study area (site=0.54, 95% confidence interval [CI]=0.34, 0.74) and was comparable to the naïve estimate (site=0.50). We did not find strong support for an effect of pH on the occurrence of ranavirus (Table 4; weight=0.17, =0.36, 95% CI=-0.90, 1.62). Ranavirus detection probability was imperfect, and varied by both species and tissue type (Fig. 2), potentially reflecting both prevalence and detection differences that could not be separated without using the multiscale model.

Table 4

Model selection results for occupancy models of ranavirus site occurrence (ψsite), prevalence (ψsite or θ), and detection probability (psite, pind, and pmult) from three analyses. The first analysis estimated ranavirus site occupancy (ψsite) and detection probability (psite) using data from 22 wetlands surveyed at the Patuxent Research Refuge (Laurel, Maryland, USA). The second analysis used a subset of the data to estimate prevalence (ψind) and detection (pind) at a single focal site where 12 spotted salamander (Ambystoma maculatum) larvae and 12 wood frog (Lithobates sylvaticus) tadpoles were sampled. The final analysis estimated site occurrence (ψsite), prevalence (θ), and detection probability (pmult) using data from all sites and sampled individuals using multiscale occupancy models. In each analysis we examined one site covariate (pH) and two detection covariates (tissue type [tissue] and species [sp]) in addition to null (.) models, and ran all possible combinations of these covariates, including an additive effect of tissue and species. For the multiscale analysis we also considered that prevalence (θ) could vary by species. The number of parameters (K), difference in Akaike's information criterion adjusted for small sample sizes (ΔAICc), model weight, and deviance are given for each model.

Model selection results for occupancy models of ranavirus site occurrence (ψsite), prevalence (ψsite or θ), and detection probability (psite, pind, and pmult) from three analyses. The first analysis estimated ranavirus site occupancy (ψsite) and detection probability (psite) using data from 22 wetlands surveyed at the Patuxent Research Refuge (Laurel, Maryland, USA). The second analysis used a subset of the data to estimate prevalence (ψind) and detection (pind) at a single focal site where 12 spotted salamander (Ambystoma maculatum) larvae and 12 wood frog (Lithobates sylvaticus) tadpoles were sampled. The final analysis estimated site occurrence (ψsite), prevalence (θ), and detection probability (pmult) using data from all sites and sampled individuals using multiscale occupancy models. In each analysis we examined one site covariate (pH) and two detection covariates (tissue type [tissue] and species [sp]) in addition to null (.) models, and ran all possible combinations of these covariates, including an additive effect of tissue and species. For the multiscale analysis we also considered that prevalence (θ) could vary by species. The number of parameters (K), difference in Akaike's information criterion adjusted for small sample sizes (ΔAICc), model weight, and deviance are given for each model.
Model selection results for occupancy models of ranavirus site occurrence (ψsite), prevalence (ψsite or θ), and detection probability (psite, pind, and pmult) from three analyses. The first analysis estimated ranavirus site occupancy (ψsite) and detection probability (psite) using data from 22 wetlands surveyed at the Patuxent Research Refuge (Laurel, Maryland, USA). The second analysis used a subset of the data to estimate prevalence (ψind) and detection (pind) at a single focal site where 12 spotted salamander (Ambystoma maculatum) larvae and 12 wood frog (Lithobates sylvaticus) tadpoles were sampled. The final analysis estimated site occurrence (ψsite), prevalence (θ), and detection probability (pmult) using data from all sites and sampled individuals using multiscale occupancy models. In each analysis we examined one site covariate (pH) and two detection covariates (tissue type [tissue] and species [sp]) in addition to null (.) models, and ran all possible combinations of these covariates, including an additive effect of tissue and species. For the multiscale analysis we also considered that prevalence (θ) could vary by species. The number of parameters (K), difference in Akaike's information criterion adjusted for small sample sizes (ΔAICc), model weight, and deviance are given for each model.
Figure 2

Estimates and 95% confidence intervals for detection probability (the probability of a PCRpositive result, given an individual is infected with ranavirus) for two tissue sample types (lethal liver sample and nonlethal tail samples) from site occurrence (gray bars) and multiscale (white bars) analyses of ranavirus occurrence at 22 wetlands in the Patuxent Research Refuge (Laurel, Maryland, USA). Ranavirus detection estimates are shown for both spotted salamanders (Ambystoma maculatum, light gray) and wood frogs (Lithobates sylvaticus, dark gray). Estimates shown are from the best-supported (i.e., lowest Akaike's information criterion adjusted for small sample sizes) site occurrence (gray) and multiscale (white) occupancy models.

Figure 2

Estimates and 95% confidence intervals for detection probability (the probability of a PCRpositive result, given an individual is infected with ranavirus) for two tissue sample types (lethal liver sample and nonlethal tail samples) from site occurrence (gray bars) and multiscale (white bars) analyses of ranavirus occurrence at 22 wetlands in the Patuxent Research Refuge (Laurel, Maryland, USA). Ranavirus detection estimates are shown for both spotted salamanders (Ambystoma maculatum, light gray) and wood frogs (Lithobates sylvaticus, dark gray). Estimates shown are from the best-supported (i.e., lowest Akaike's information criterion adjusted for small sample sizes) site occurrence (gray) and multiscale (white) occupancy models.

Prevalence

At the single focal site, naïve estimates of ranavirus prevalence were 0.75 for wood frogs and 0.17 for spotted salamanders. Model selection results from the prevalence analysis supported the idea that prevalence differed by species (Table 4), but estimates of prevalence were considerably higher than the naïve values. At this single site, ranavirus was estimated to be ubiquitous in wood frogs (ind,frog=1, SE=0) and was less prevalent in spotted salamanders (ind,salamander=0.26, 95% CI=0.06,0.68). In addition, the best-supported model suggested that detection probability was constant, and did not vary by tissue type or species at this site (Table 4; ind=0.29, 95% CI=0.17, 0.44).

Multiscale: occupancy and prevalence

Without accounting for detection probability, the naïve estimates of prevalence across the study area were identical for both species (42%; Fig. 3). However, the best-supported multiscale occupancy model produced prevalence (θ) estimates that differed by species and were much higher than naïve estimates (Table 4 and Fig. 3). In accordance with the site occurrence analysis presented, the multiscale analysis produced an occurrence estimate that was very similar to that estimated using the standard occupancy model (site=0.55, 95% CI=0.35, 0.74) and found no effect of pH on ranavirus occurrence (Table 4). Ranavirus detection probability in tail samples was higher than in liver samples (mult,liver=0.60, 95% CI=0.52, 0.68, mult,tail= 0.91, 95% CI=0.87, 0.94; Fig. 2). We provide a tutorial (Supplementary Materials) that readers can use to perform the analyses we present in Program PRESENCE.

Figure 3

Naïve (dark gray) and estimated (light gray) ranavirus prevalence with 95% confidence intervals for spotted salamander (Ambystoma maculatum) and wood frog (Lithobates sylvaticus) larvae at 22 wetlands in the Patuxent Research Refuge Refuge (Laurel, Maryland, USA). Estimates shown are from the best-supported multiscale occupancy model (i.e., lowest Akaike's information criterion adjusted for small sample sizes). Confidence intervals are not provided for naïve estimates because the binomial variance associated with these estimates does not account for detection uncertainty, and is therefore overly precise.

Figure 3

Naïve (dark gray) and estimated (light gray) ranavirus prevalence with 95% confidence intervals for spotted salamander (Ambystoma maculatum) and wood frog (Lithobates sylvaticus) larvae at 22 wetlands in the Patuxent Research Refuge Refuge (Laurel, Maryland, USA). Estimates shown are from the best-supported multiscale occupancy model (i.e., lowest Akaike's information criterion adjusted for small sample sizes). Confidence intervals are not provided for naïve estimates because the binomial variance associated with these estimates does not account for detection uncertainty, and is therefore overly precise.

DISCUSSION

Detecting pathogens in the environment and understanding among-site differences in the proportion of infected hosts are primary goals in surveillance studies (Richgels et al. 2016; Gray et al. 2017), and the consequences of underestimating pathogen occurrence can be large. While studies frequently acknowledge that pathogen detection is imperfect, probabilistic approaches to dealing with this uncertainty are rarely used. We demonstrated the utility of occupancy models for addressing pathogen occurrence questions using readily available data from an easily employed sampling design.

Liver is the diagnostic tissue of choice for detecting ranavirus infection, especially in larval amphibians (Miller et al. 2015). Gray et al. (2012) reported a higher detection rate for ranavirus in liver samples compared to swabs and tail clips, but assumed that liver tissue was a perfect diagnostic sample (p=1) and that any detection in a tail sample when the liver tested negative was a false positive. False positives could occur due to external virions from the environment adhering to the amphibian's skin (Gray et al. 2012). However, it is also possible that early stages of infection can be detected in vascularized tail samples as the animal becomes viremic (Miller et al. 2015). When we detected ranavirus in a liver sample, there was a 94% probability of subsequent detection in at least one tail sample, suggesting that tail samples may be good indicators of host infection (i.e., diagnostic of true disease state, and not simply pathogen presence). Differing detection probabilities between sample types are likely to converge as ranavirus infections become systemic (Greer and Collins 2007), but given the value of detecting ranavirus before population declines occur, early detection at a site may be desirable even if not directly linked to host infections. Nonlethal tail samples are a reasonable alternative to lethal sampling for ranavirus, especially when lethal means are prohibited or when site-level occurrence dynamics are the focal question. Using a more sensitive detection method also has important study design benefits in that fewer animals need to be detected, captured, and sampled for robust inferences.

Detection probabilities were imperfect (<1) across all models, and the best-supported models suggested detection variation by tissue type. Thus, pathogen prevalence was underestimated when imperfect detection was not accounted for, as expected (Greer and Collins 2007; McClintock et al. 2010; Lachish et al. 2012). The detection probability estimated by the site occurrence model set was lower, and had a different best-supported structure (i.e., lowest AICc), than that estimated using the same data in the multiscale model. This difference occurred because the estimate of p from the basic occupancy model was the product of prevalence and detection probability (i.e., the positive predictive value). This was because the standard site analysis of occurrence violated the closure assumption and assumed that there was an opportunity for pathogen detection in every tissue sample (though some individuals may not have been infected, eliminating the possibility of detection). The multiscale model accounted for this violation by allowing the subunits sampled (individuals) to have different disease states. Without this accommodation, detection estimates will be biased low and occurrence estimates will be biased high (Kendall and White 2009).

Identifying the variables that shape ranavirus disease dynamics is critical for developing effective in situ management actions to cope with amphibian die-offs. While we found no relationship between pH and ranavirus occurrence, there was a narrow range of pH values at the PRR (range=4.43–6.96). Future studies should explore additional habitat characteristics predicted to influence pathogen occurrence and disease dynamics, including host community composition, abiotic characteristics such as contaminants (including endocrine-disrupting compounds), wetland permanence, or the effect of habitat manipulations. The occupancy modeling framework we present yielded unbiased estimates of covariate effects on pathogen occurrence and prevalence, and can be used to identify important drivers of disease dynamics.

Within ranavirus-positive wetlands, the ranavirus prevalence differed by species and was similar to the naïve prevalence estimates in controlled experiments for these species (Hoverman et al. 2011). Therefore, susceptibility trials might provide reasonable insights into field epidemiology (Gray et al. 2017) if imperfect detection is accounted for. Prevalence alone is not necessarily an indication of the degree of morbidity expected, because species may be differentially susceptible to infection (Hoverman et al. 2011), and may have variable infection loads (Hall et al. 2016). However, because wood frogs are known to have a strong correlation between infection and mortality (Hoverman et al. 2011), pathogen prevalence might be a reasonable estimate of disease impacts for this species. Spotted salamander larvae are more tolerant to ranavirus infection (Hoverman et al. 2011), meaning that they suffer fewer fitness consequences from ranavirus infection than other tested amphibian species. Prolonged ranavirus infections could make spotted salamanders biologic reservoirs that contribute to the emergence (and reemergence) of ranavirus in coexisting wood frog tadpole populations. Our estimates of ranavirus prevalence were higher than those in some field studies (Crespi et al. 2015), likely because other studies have not accounted for detection probability. While failing to account for heterogeneity in detection makes comparisons across landscapes and among studies difficult, the adoption of tools presented in this manuscript will facilitate future comparisons, which can be accomplished by accounting for study-specific sampling error.

An occupancy-based study design frees the researcher from the need to sample during disease-induced die-offs, as is recommended in some other designs (Simon and Schill 1984; Lavilla-Pitogo and de la Pen˜ a 2004). While pathogen detection may be maximized during a die-off, resulting data offer few insights about enzootic dynamics or spatiotemporal variation in prevalence (Greer et al. 2009; Hoverman et al. 2012). In addition, there is a risk of missing a die-off completely, especially in systems where mortality can occur rapidly, such as with ranavirus (Wheelwright et al. 2014; Hall et al. 2016), leaving uncertainty about the presence of a population or the influence of disease on persistence. Instead, an occupancy framework allows for a wider sampling window because detection probability is explicitly modeled and may be allowed to vary with time and sampling method. Estimates of detection probability are likely related to infection intensity (Miller et al. 2012) and therefore may differ based on sampling conditions like time until metamorphosis, time of year, or time since ranavirus invasion. However, the occupancy framework allows researchers to investigate the influences of these conditions, provided the study design is adequate (i.e., with a longitudinal study).

One should consider the tradeoffs between sampling effort, cost, and timing when designing studies to address questions about pathogen occurrence. Occupancy-based findings from a study in a single year can be used as pilot data to inform the timing, location, and sample sizes for future sampling efforts, potentially conserving or optimizing resource expenditures. Given the prevalence of ranavirus in woods frogs at the PRR and the tail-clip detection probability, we can be 95% confident in identifying ranavirus at least once at a site by collecting single tail clip samples from just three individuals rather than the 12 that were part of our initial study design (confidence =1-[1-θp]samples).

In our study we focused on estimating pathogen occurrence at a single point in time. However, static snapshots of dynamic systems (i.e., cross-sectional studies) can sometimes be misleading (Yackulic et al. 2015). We recommend that the sampling designs presented here be incorporated into monitoring efforts (i.e., longitudinal studies) so that pathogen dynamics (i.e., annual colonization and extinction of sites, changes in infection prevalence or incidence) can be estimated and used in conservation decision-making. The occupancy models we presented can be extended to include multiple seasons (MacKenzie et al. 2003), false positive errors (Miller et al. 2011), and multiple host-pathogen states (Mosher et al. 2018).

When hosts are scarce, detection of shed ranavirus DNA is possible using eDNA methods (Hall et al. 2016). Standard and multiscale occupancy models are able to accommodate eDNA data (Schmidt et al. 2013; Mosher et al. 2017), and the framework and tutorial we present here can be used as a guide for making unbiased inference from these kinds of data.

ACKNOWLEDGMENTS

The study was supported by the US Geological Survey's Amphibian Research and Monitoring Initiative (ARMI), the US Fish and Wildlife Service Region 5, and the US Department of Agriculture National Institute of Food and Agriculture (Hatch project 1012932). This is contribution number 669 of the ARMI program. We thank the Northeast Amphibian Research and Monitoring Initiative field crew for collecting field data. The research was approved by the Patuxent Research Refuge Animal Care and Use Committee. Any use of trade, firm, or product name is for descriptive purposes only and does not imply endorsement by the US Government.

SUPPLEMENTARY MATERIAL

Supplementary material for this article is online at http://dx.doi.org/10.7589/2018-02-042.

LITERATURE CITED

LITERATURE CITED
Adams
MJ
,
Chelgren
ND
,
Reinitz
D
,
Cole
RA
,
Rachowicz
LJ
,
Galvan
S
,
McCreary
B
,
Pearl
CA
,
Bailey
LL
,
Bettaso
J
, et al.
2010
.
Using occupancy models to understand the distribution of an amphibian pathogen, Batrachochytrium dendrobatidis.
Ecol Appl
20
:
289
302
.
Brunner
JL
,
Storfer
A
,
Gray
MJ
,
Hoverman
JT
.
2015
.
Ranavirus ecology and evolution: From epidemiology to extinction.
In
:
Ranaviruses
,
Gray
MJ
,
Chinchar
VG
, editors.
Springer
,
Cham, Switzerland
, pp.
71
104
.
Burnham
KP
,
Anderson
DR
.
2003
.
Model selection and multimodel inference: A practical information-theoretic approach.
Springer
,
New York, New York
,
488
pp.
Caley
P
,
Ramsey
D
.
2001
.
Estimating disease transmission in wildlife, with emphasis on leptospirosis and bovine tuberculosis in possums, and effects of fertility control.
J Appl Ecol
38
:
1362
1370
.
Chinchar
VG
.
2002
.
Ranaviruses (family Iridoviridae): Emerging cold-blooded killers.
Arch Virol
147
:
447
470
.
Crespi
EJ
,
Rissler
LJ
,
Mattheus
NM
,
Engbrecht
K
,
Duncan
SI
,
Seaborn
T
,
Hall
EM
,
Peterson
JD
,
Brunner
JL
.
2015
.
Geophysiology of wood frogs: Landscape patterns of prevalence of disease and circulating hormone concentrations across the eastern range.
Integr Comp Biol
55
:
602
617
.
Daszak
P
,
Cunningham
AA
,
Hyatt
AD
.
2000
.
Emerging infectious diseases of wildlife—Threats to biodiversity and human health.
Science
287
:
443
449
.
Daszak
P
,
Cunningham
AA
,
Hyatt
AD
.
2003
.
Infectious disease and amphibian population declines.
Divers Distrib
9
:
141
150
.
DiRenzo
GV
,
Campbell
Grant EH
,
Longo
AV
,
Che-Castaldo
C
,
Zamudio
KR
,
Lips
KR
.
2018
.
Imperfect pathogen detection from non-invasive skin swabs biases disease inference.
Methods Ecol Evol
9
:
380
389
.
Duffus
AL
,
Waltzek
TB
,
Stöhr
AC
,
Allender
MC
,
Gotesman
M
,
Whittington
RJ
,
Hick
P
,
Hines
MK
,
Marschang
RE
.
2015
.
Distribution and host range of ranaviruses.
In
:
Ranaviruses
,
Gray
MJ
Chinchar
VG
, editors.
Springer
,
Cham, Switzerland
, pp.
9
57
.
Elmore
SA
,
Huyvaert
KP
,
Bailey
LL
,
Iqbal
A
,
Su
C
,
Dixon
BR
,
Alisauskas
RT
,
Gajadhar
AA
,
Jenkins
EJ
.
2016
.
Multi-scale occupancy approach to estimate Toxoplasma gondii prevalence and detection probability in tissues: An application and guide for field sampling.
Int J Parasitol
46
:
563
570
.
Elmore
SA
,
Huyvaert
KP
,
Bailey
LL
,
Milhous
J
,
Alisauskas
RT
,
Gajadhar
AA
,
Jenkins
EJ
.
2014
.
Toxoplasma gondii exposure in arctic-nesting geese: A multi-state occupancy framework and comparison of serological assays.
Int J Parasitol Parasites Wildl
3
:
147
153
.
Fisher
MC
,
Henk
DA
,
Briggs
CJ
,
Brownstein
JS
,
Madoff
LC
,
McCraw
SL
,
Gurr
SJ
.
2012
.
Emerging fungal threats to animal, plant and ecosystem health.
Nature
484
:
186
194
.
Gerber
BD
,
Converse
SJ
,
Muths
E
,
Crockett
HJ
,
Mosher
BA
,
Bailey
LL
.
2018
.
Identifying species conservation strategies to reduce disease-associated declines.
Conserv Lett
11
:
e12393
.
Goldberg
CS
,
Turner
CR
,
Deiner
K
,
Klymus
KE
,
Thomsen
PF
,
Murphy
MA
,
Spear
SF
,
McKee
A
,
Oyler-McCance
SJ
,
Cornman
RS
, et al.
2016
.
Critical considerations for the application of environmental DNA methods to detect aquatic species.
Methods Ecol Evol
7
:
1299
1307
.
Grant
EHC
,
Muths
E
,
Katz
RA
,
Canessa
S
,
Adams
MJ
,
Ballard
JR
,
Berger
L
,
Briggs
CJ
,
Coleman
JT
,
Gray
MJ
, et al.
2017
.
Using decision analysis to support proactive management of emerging infectious wildlife diseases.
Front Ecol Environ
15
:
214
221
.
Gray
MJ
,
Duffus
ALJ
,
Haman
KH
,
Harris
RN
,
Allender
MC
,
Thompson
TA
,
Christman
MR
,
Sacerdote-Velat
A
,
Sprague
LA
,
Williams
JM
, et al.
2017
.
Pathogen surveillance in herpetofaunal populations: Guidance on study design, sample collection, biosecurity, and intervention strategies.
Herpetol Rev
48
:
334
351
.
Gray
MJ
,
Miller
DL
,
Hoverman
JT
.
2012
.
Reliability of non-lethal surveillance methods for detecting ranavirus infection.
Dis Aquat Organ
99
:
1
6
.
Green
DE
,
Converse
KA
,
Schrader
AK
.
2002
.
Epizootiology of sixty-four amphibian morbidity and mortality events in the USA, 1996–2001.
Ann N Y Acad Sci
969
:
323
339
.
Green
DE
,
Gray
MJ
,
Miller
DL
.
2009
.
Disease monitoring and biosecurity.
In
:
Amphibian ecology and conservation: A handbook of techniques
,
Dodd
C
, editor.
Oxford University Press
,
Oxford, UK
, pp.
481
505
.
Greer
AL
,
Brunner
JL
,
Collins
JP
.
2009
.
Spatial and temporal patterns of Ambystoma tigrinum virus (ATV) prevalence in tiger salamanders Ambystoma tigrinum nebulosum.
Dis Aquat Organ
85
:
1
6
.
Greer
AL
,
Collins
JP
.
2007
.
Sensitivity of a diagnostic test for amphibian ranavirus varies with sampling protocol.
J Wildl Dis
43
:
525
532
.
Hall
EM
,
Crespi
EJ
,
Goldberg
CS
,
Brunner
JL
.
2016
.
Evaluating environmental DNA-based quantification of ranavirus infection in wood frog populations.
Mol Ecol Resour
16
:
423
433
.
Hines
JE.
2006
.
Software to generate patch occupancy data and analyze using programs MARK or PRESENCE.
Accessed February 2018
Hoverman
JT
,
Gray
MJ
,
Haislip
NA
,
Miller
DL
.
2011
.
Phylogeny, life history, and ecology contribute to differences in amphibian susceptibility to ranaviruses.
Ecohealth
8
:
301
319
.
Hoverman
JT
,
Gray
MJ
,
Miller
DL
,
Haislip
NA
.
2012
.
Widespread occurrence of ranavirus in pond-breeding amphibian populations.
Ecohealth
9
:
36
48
.
Jennelle
CS
,
Cooch
EG
,
Conroy
MJ
,
Senar
JC
.
2007
.
State-specific detection probabilities and disease prevalence.
Ecol Appl
17
:
154
167
.
Johnson
PT
,
De
Roode JC
,
Fenton
A
.
2015
.
Why infectious disease research needs community ecology.
Science
349
:
1259504
.
Keesing
F
,
Holt
RD
,
Ostfeld
RS
.
2006
.
Effects of species diversity on disease risk.
Ecol Lett
9
:
485
498
.
Kendall
WL
,
White
GC
.
2009
.
A cautionary note on substituting spatial subunits for repeated temporal sampling in studies of site occupancy.
J Appl Ecol
46
:
1182
1188
.
Lachish
S
,
Gopalaswamy
AM
,
Knowles
SCL
,
Sheldon
BC
.
2012
.
Site-occupancy modelling as a novel framework for assessing test sensitivity and estimating wildlife disease prevalence from imperfect diagnostic tests.
Methods Ecol Evol
3
:
339
348
.
Langwig
KE
,
Voyles
J
,
Wilber
MQ
,
Frick
WF
,
Murray
KA
,
Bolker
BM
,
Collins
JP
,
Cheng
TL
,
Fisher
MC
,
Hoyt
JR
, et al.
2015
.
Context-dependent conservation responses to emerging wildlife diseases.
Front Ecol Environ
13
:
195
202
.
Lavilla-Pitogo
CR
,
de
la Pen˜ a LD
.
2004
.
Diseases in farmed mud crabs Scylla spp.: Diagnosis, prevention, and control.
Aquaculture Department
,
Southeast Asian Fisheries Development Center, Tigbauan, Iloilo, Philippines
,
89
pp.
Lorch
JM
,
Muller
LK
,
Russell
RE
,
O'Connor
M
,
Lindner
DL
,
Blehert
DS
.
2013
.
Distribution and environmental persistence of the causative agent of whitenose syndrome, Geomyces destructans, in bat hibernacula of the eastern United States.
Appl Environ Microbiol
79
:
1293
1301
.
MacKenzie
DI
,
Nichols
JD
,
Hines
JE
,
Knutson
MG
,
Franklin
AB
.
2003
.
Estimating site occupancy, colonization, and local extinction when a species is detected imperfectly.
Ecology
84
:
2200
2207
.
MacKenzie
DI
,
Nichols
JD
,
Lachman
GB
,
Droege
S
,
Andrew
Royle J
,
Langtimm
CA
.
2002
.
Estimating site occupancy rates when detection probabilities are less than one.
Ecology
83
:
2248
2255
.
MacKenzie
DI
,
Nichols
JD
,
Royle
JA
,
Pollock
KH
,
Bailey
L
,
Hines
JE
.
2018
.
Occupancy estimation and modeling: Inferring patterns and dynamics of species occurrence. 2nd Ed.
Academic Press
,
San Diego, California
,
648
pp.
McClintock
BT
,
Nichols
JD
,
Bailey
LL
,
MacKenzie
DI
,
Kendall
WL
,
Franklin
AB
.
2010
.
Seeking a second opinion: Uncertainty in disease ecology.
Ecol Lett
13
:
659
674
.
McKee
AM
,
Spear
SF
,
Pierson
TW
.
2015
.
The effect of dilution and the use of a post-extraction nucleic acid purification column on the accuracy, precision, and inhibition of environmental DNA samples.
Biol Conserv
183
:
70
76
.
Miller
DA
,
Nichols
JD
,
McClintock
BT
,
Grant
EHC
,
Bailey
LL
,
Weir
LA
.
2011
.
Improving occupancy estimation when two types of observational error occur: Non-detection and species misidentification.
Ecology
92
:
1422
1428
.
Miller
DAW
,
Talley
BL
,
Lips
KR
,
Campbell
Grant EH
.
2012
.
Estimating patterns and drivers of infection prevalence and intensity when detection is imperfect and sampling error occurs.
Methods Ecol Evol
3
:
850
859
.
Miller
DL
,
Pessier
AP
,
Hick
P
,
Whittington
RJ
.
2015
.
Comparative pathology of ranaviruses and diagnostic techniques.
In:
Ranaviruses,
Gray
MJ
,
Chinchar
VG
, editors.
Springer
,
Cham, Switzerland
, pp.
171
208
.
Mosher
BA
,
Bailey
LL
,
Hubbard
BA
,
Huyvaert
KP
.
2018
.
Inferential biases linked to unobservable states in complex occupancy models.
Ecography
41
:
32
39
.
Mosher
BA
,
Huyvaert
KP
,
Chestnut
T
,
Kerby
JL
,
Madison
JD
,
Bailey
LL
.
2017
.
Design- and modelbased strategies for detecting and quantifying an amphibian pathogen in environmental samples.
Ecol Evol
7
:
10952
10962
.
Muths
E
,
Bailey
LL
,
Watry
MK
.
2014
.
Animal reintroductions: An innovative assessment of survival.
Biol Conserv
172
:
200
208
.
Nichols
JD
,
Bailey
LL
,
O'Connell
AF Jr
,
Talancy
NW
,
Campbell
Grant EH
,
Gilbert
AT
,
Annand
EM
,
Husband
TP
,
Hines
JE
.
2008
.
Multi-scale occupancy estimation and modelling using multiple detection methods.
J Appl Ecol
45
:
1321
1329
.
Price
SJ
,
Garner
TW
,
Nichols
RA
,
Balloux
F
,
Ayres
C
,
de
Alba AM-C
,
Bosch
J
.
2014
.
Collapse of amphibian communities due to an introduced Ranavirus.
Curr Biol
24
:
2586
2591
.
Rachowicz
LJ
,
Hero
J
,
Alford
RA
,
Taylor
JW
,
Morgan
JAT
,
Vredenburg
VT
,
Collins
JP
,
Briggs
CJ
.
2005
.
The novel and endemic pathogen hypotheses: Competing explanations for the origin of emerging infectious diseases of wildlife.
Conserv Biol
19
:
1441
1448
.
Richgels
KL
,
Russell
RE
,
Adams
MJ
,
White
CL
,
Grant
EHC
.
2016
.
Spatial variation in risk and consequence of Batrachochytrium salamandrivorans introduction in the USA.
R Soc Open Sci
3
:
150616
.
Rödder
D
,
Kielgast
J
,
Bielby
J
,
Schmidtlein
S
,
Bosch
J
,
Garner
TW
,
Veith
M
,
Walker
S
,
Fisher
MC
,
Lötters
S
.
2009
.
Global amphibian extinction risk assessment for the panzootic chytrid fungus.
Diversity
1
:
52
66
.
Rogers
DJ
,
Randolph
SE
.
2000
.
The global spread of malaria in a future, warmer world.
Science
289
:
1763
1766
.
Russell
RE
,
Katz
RA
,
Richgels
KLD
,
Walsh
DP
,
Grant
EHC
.
2017
.
A framework for modeling emerging diseases to inform management.
Emerg Infect Dis
23
:
1
6
.
Scheele
BC
,
Foster
CN
,
Banks
SC
,
Lindenmayer
DB
.
2017
.
Niche contractions in declining species: Mechanisms and consequences.
Trends Ecol Evol
32
:
346
355
.
Schmidt
BR
.
2010
.
Estimating the impact of disease in species threatened by amphibian chytrid fungus: Comment on Murray et al.
Conserv Biol
24
:
897
899
.
Schmidt
BR
,
Kéry
M
,
Ursenbacher
S
,
Hyman
OJ
,
Collins
JP
.
2013
.
Site occupancy models in the analysis of environmental DNA presence/absence surveys: A case study of an emerging amphibian pathogen.
Methods Ecol Evol
4
:
646
653
.
Simon
RC
,
Schill
WB
.
1984
.
Tables of sample size requirements for detection of fish infected by pathogens: Three confidence levels for different infection prevalence and various population sizes.
J Fish Dis
7
:
515
520
.
Sparling
DW
.
2003
.
A review of the role of contaminants in amphibian declines.
In:
Handbook of ecotoxicology, 2nd Ed
.,
Hoffman
DJ
,
Burton
GA Jr
,
Cairns
J Jr
,
Rattner
BA
, editors.
Lewis Publishers
,
Boca Raton, Florida
, pp.
1099
1128
.
Sutton
WB
,
Gray
MJ
,
Hardman
RH
,
Wilkes
RP
,
Kouba
AJ
,
Miller
DL
.
2014
.
High susceptibility of the endangered dusky gopher frog to ranavirus.
Dis Aquat Organ
112
:
9
16
.
Sutton
WB
,
Gray
MJ
,
Hoverman
JT
,
Secrist
RG
,
Super
PE
,
Hardman
RH
,
Tucker
JL
,
Miller
DL
.
2015
.
Trends in ranavirus prevalence among plethodontid salamanders in the Great Smoky Mountains National Park.
Ecohealth
12
:
320
329
.
Thompson
KG
.
2007
.
Use of site occupancy models to estimate prevalence of Myxobolus cerebralis infection in trout.
J Aquat Anim Health
19
:
8
13
.
Tyre
AJ
,
Tenhumberg
B
,
Field
SA
,
Niejalke
D
,
Parris
K
,
Possingham
HP
.
2003
.
Improving precision and reducing bias in biological surveys: Estimating falsenegative error rates.
Ecol Appl
13
:
1790
1801
.
Valenzuela-Sánchez
A
,
Schmidt
BR
,
Uribe-Rivera
DE
,
Costas
F
,
Cunningham
AA
,
Soto-Azat
C
.
2017
.
Cryptic disease-induced mortality may cause host extinction in an apparently stable host-parasite system.
Proc Biol Sci
284
:
20171176
.
Van
Meter R
,
Bailey
LL
,
Grant
EHC
.
2008
.
Methods for estimating the amount of vernal pool habitat in the northeastern United States.
Wetlands
28
:
585
593
.
Wasserberg
G
,
Osnas
EE
,
Rolley
RE
,
Samuel
MD
.
2009
.
Host culling as an adaptive management tool for chronic wasting disease in white-tailed deer: A modelling study.
J Appl Ecol
46
:
457
466
.
Wheelwright
NT
,
Gray
MJ
,
Hill
RD
,
Miller
DL
.
2014
.
Sudden mass die-off of a large population of wood frog (Lithobates sylvaticus) tadpoles in Maine, USA, likely due to ranavirus.
Herpetol Rev
45
:
240
242
.
Woodhams
DC
,
Bosch
J
,
Briggs
CJ
,
Cashins
S
,
Davis
LR
,
Lauer
A
,
Muths
E
,
Puschendorf
R
,
Schmidt
BR
,
Sheafor
B
, et al.
2011
.
Mitigating amphibian disease: Strategies to maintain wild populations and control chytridiomycosis.
Front Zool
8
:
8
.
Yackulic
CB
,
Nichols
JD
,
Reid
J
,
Der
R
.
2015
.
To predict the niche, model colonization and extinction.
Ecology
96
:
16
23
.

Supplementary data