Site‐occupancy modelling as a novel framework for assessing test sensitivity and estimating wildlife disease prevalence from imperfect diagnostic tests

Abstract

Summary1. Reliable assessments of infection status and population prevalence are critical for epidemiological modelling and disease management, but can be greatly biased when disease state is determined from imperfect diagnostic tests. Available statistical methods to adjust test-based prevalence estimates by correcting for test accuracy demand that many stringent requirements and assumptions be met (knowledge about underlying population prevalence or multiple diagnostic methods), limiting their utility for wildlife disease surveys. 2. In this paper, we present site-occupancy modelling as a flexible approach to derive estimates of population prevalence and test sensitivity under imperfect pathogen detection without a need for restrictive requirements or assumptions. We extend the utility of the standard site-occupancy framework for pathogen detection data by novel application of abundance-induced heterogeneity (AIH) models (Royle & Nichols 2003) that allow test sensitivity to vary with host pathogen load or infection intensity. 3. We demonstrate the utility of this approach for wildlife disease studies by applying site-occupancy models to a data set consisting of replicate quantitative (q)PCR diagnoses of malaria parasites (Plasmodium spp.) in blood samples from wild blue tits (Cyanistes caeruleus). 4. Model selection revealed that Plasmodium detection rates by qPCR were strongly dependent on host parasite load. Estimates of parasite detection rates revealed the qPCR assay to be highly sensitive, with accordingly, a very low probability of false negative diagnosis for the majority of infected hosts in our population and little bias in naive estimates of population prevalence, although this will be a system-specific result. 5. Our results demonstrate the utility of a site-occupancy approach for deriving estimates of population prevalence under imperfect pathogen detection and reveal that accounting for host variation in pathogen load allows a more accurate assessment of diagnostic test sensitivity. 6. By identifying factors that influence pathogen detection rates, and revealing optimal protocols for obtaining unbiased prevalence estimates, while minimising the probability of false negative diagnoses, we also show that this approach can enhance both diagnostic accuracy and cost-efficiency in wildlife disease surveys.

Spatial determinants of infection risk in a multi‐species avian malaria system

“…From 2005 to 2010 blood samples for infection diagnosis were collected from individually marked blue tits between day 6 and 14 of the nestling phase. Analyses using occupancy modelling have shown the qPCR assay technique is highly sensitive, with a very low probability of false negative diagnoses (Lachish et al 2012); strict laboratory protocols ensure the likelihood of false positive diagnoses is small (Knowles et al 2011). As these populations are single brooded and breeding is synchronous, there is relatively little variation in the calendar date among samples within each year (average range Ϯ SE ϭ 42.42 Ϯ 6.78 d).…”

Modeling false positive detections in species occurrence data under different study designs

“…This 161 independent assessment is performed in reference sites where the true occupancy state is 162 controlled by the investigator and therefore known with certainty. Such data are very common in 163 lab studies (e.g., eco-epidemiology, eco-immunology) based on molecular assays for the 164 detection of specific DNA fragments, antigens or circulating antibodies (Lachish et al 2012). The sampling design is identical to the one presented in the previous section for an 175 ambiguous detection method, where detection/non-detection data are collected ( ) in which 176 both site-level false positives and false negatives can occur.…”

Multilevel Models for the Distribution of Hosts and Symbionts

“…Continuous distributions would circumvent the need to round values for use with Poisson or negative binomial distributions with integer support. Last, we have assumed that infections are detected without error, but a rich set of methods could be applied to account for error in this measurement process [45, 46]. …”