Imperfect pathogen detection from non‐invasive skin swabs biases disease inference

Abstract

Conservation managers rely on accurate estimates of disease parameters, such as pathogen prevalence and infection intensity, to assess disease status of a host population. However, these disease metrics may be biased if low‐level infection intensities are missed by sampling methods or laboratory diagnostic tests. These false negatives underestimate pathogen prevalence and overestimate mean infection intensity of infected individuals. Our objectives were two‐fold. First, we quantified false negative error rates of Batrachochytrium dendrobatidis (Bd) on non‐invasive skin swabs collected from an amphibian community in El Copé, Panama. We swabbed amphibians twice in sequence, and we used a recently developed hierarchical Bayesian estimator to assess disease status of the population. Second, we developed a novel hierarchical Bayesian model to simultaneously account for imperfect pathogen detection from field sampling and laboratory diagnostic testing. We evaluated the performance of the model, using simulations and varying sampling design to quantify the magnitude of bias in estimates of pathogen prevalence and infection intensity. We show that Bd detection probability from skin swabs was related to host infection intensity, where Bd infections <10 zoospores have <95% probability of being detected. If imperfect Bd detection was not considered, then Bd prevalence was underestimated by as much as 71%. In the Bd‐amphibian system, this indicates a need to correct for imperfect pathogen detection in enzootic host populations persisting with low‐level infections. More generally, our results have implications for study designs in other disease systems, particularly those with similar objectives, biology, and sampling decisions. Uncertainty in pathogen detection is an inherent property of most sampling protocols and diagnostic tests, where the magnitude of bias depends on the study system, type of infection, and false negative error rates. Given that it may be difficult to know this information in advance, we advocate that the most cautious approach is to assume all errors are possible and to accommodate them by adjusting sampling designs. The modelling framework presented here improves the accuracy in estimating pathogen prevalence and infection intensity.

Disease‐structuredN‐mixture models: A practical guide to model disease dynamics using count data

“…We specify the observation model to account for imperfect host and pathogen detection during the sampling process by assuming that pathogen detection is related to infection intensities (e.g.,DiRenzo, Campbell Grant, et al, ; Lachish et al, ; Miller et al, ). We denote g s,i,j as the number of hosts detected in each disease state s (where s = 1 for uninfected and s = 2 for infected) at site i during survey replicate j.…”

“…Obtaining inference for disease dynamics is challenging because (a) demographic rates (e.g., survival, transmission; Table ) are difficult to quantify if either population size or recapture rates are low and (b) there are multiple ways in which sampling error affects the observed data (Figure ). For example, sampling error relevant to disease studies can manifest in two forms: (a) uncertainty of host occurrence or abundance (i.e., imperfect host detection), and (b) uncertainty of pathogen occurrence or abundance (i.e., imperfect pathogen detection; DiRenzo, Campbell Grant, et al, ; Lachish, Gopalaswamy, Knowles, & Sheldon, ; Miller, Talley, Lips, & Grant, ). Historically, sampling error has been either ignored or acknowledged but not modeled (reviewed by McClintock et al, ).…”

“…This misclassification can lead to biases in estimates of state‐specific survival and transition probabilities and lowers the power to detect differences between survival probabilities of disease states. Across multiple host–pathogen systems, pathogen detection probability using routine laboratory assays shows positive correlations with host infection intensity, such as Batrachochytrium dendrobatidis ( Bd ) zoospores on amphibian skin (DiRenzo, Campbell Grant, et al, ; Miller et al, ); the causative agent of malaria, Plasmodium sp., in birds (Knowles et al, ; Lachish et al, ); the causative agent of tuberculosis, Mycobacterium bovis , in cattle (Drewe, Dean, Michel, & Pearce, ); and the causative agent of Lyme disease, the bacterium Borrelia , in Ixodes uriae ticks (Gómez‐Díaz, Doherty, Duneau, & McCoy, ). This suggests that infected individuals with low infection intensities are likely to be misclassified as uninfected in many systems using a variety of methods.…”

“…Each observed host is assigned a disease state (infected or not infected), based on diagnostic symptoms of disease or laboratory analysis (e.g., swabs, blood, or tissue samples), and used to calculate the average infection intensity at site i during survey j . We assume that a fraction of infected hosts are misidentified as uninfected because of low pathogen infection intensities (e.g., DiRenzo, Campbell Grant, et al, ; Lachish et al, ; Miller et al, ), but misidentification may also be unrelated to infection intensity and attributable to other factors (e.g., field sampling and laboratory diagnostic testing).…”

“…In cases where such data are unavailable, information obtained from outside studies can be used. The relationship between pathogen intensity and detection probability has been established for many disease systems (Drewe et al, ; Gómez‐Díaz et al, ; Knowles et al, ; Lachish et al, ; Miller et al, , DiRenzo, Campbell Grant, et al, ), which can be readily translated into informative priors (e.g., DiRenzo, Zipkin, et al, ).…”