Effectiveness of joint species distribution models in the presence of imperfect detection

Abstract

Joint species distribution models (JSDMs) are a recent development in biogeography and enable the spatial modelling of multiple species and their interactions and dependencies. However, most models do not consider imperfect detection, which can significantly bias estimates. This is one of the first papers to account for imperfect detection when fitting data with JSDMs and to explore the complications that may arise. A multivariate probit JSDM that explicitly accounts for imperfect detection is proposed, and implemented using a Bayesian hierarchical approach. We investigate the performance of the JSDM in the presence of imperfect detection for a range of factors, including varied levels of detection and species occupancy, and varied numbers of survey sites and replications. To understand how effective this JSDM is in practice, we also compare results to those from a JSDM that does not explicitly model detection but instead makes use of ‘collapsed data’. A case study of owls and gliders in Victoria, Australia, is also illustrated. Using simulations, we found that the JSDMs explicitly accounting for detection can accurately estimate intrinsic correlation between species with enough survey sites and replications. Reducing the number of survey sites decreases the precision of estimates, while reducing the number of survey replications can lead to biased estimates. For low probabilities of detection, the model may require a large number of survey replications to remove bias from estimates. However, JSDMs not explicitly accounting for detection may have a limited ability to disentangle detection from occupancy, which substantially reduces their ability to accurately infer the species distribution spatially. Our case study showed positive correlation between Sooty Owls and Greater Gliders, despite a low number of survey replications. To avoid biased estimates of inter‐species correlations and species distributions, imperfect detection needs to be considered. However, for low probability of detection, the JSDMs explicitly accounting for detection is data hungry. Estimates from such models may still be subject to bias. To overcome the bias, researchers need to carefully design surveys and choose appropriate modelling approaches. The survey design should ensure sufficient survey replications for unbiased inferences on species inter‐dependencies and occupancy.

A unified framework for time‐to‐detection occupancy and abundance models

“…This extension has been addressed in previous works (Bornand et al., 2014; Garrard et al., 2008; Halstead et al., 2018; Priyadarshani et al., 2022), making it a straightforward addition to our models. While our study primarily focused on single‐species single‐season TTD modelling, these models can also be expanded to handle multi‐species cases (Garrard et al., 2013), thereby improving estimation effectiveness (Hogg et al., 2021) and revealing species interactions. As observed in the Karoo bird data analysis, certain species produce similar estimates for the aggregation index and community parameter.…”

Joint species distribution models with imperfect detection for high‐dimensional spatial data

“…Many JSDMs jointly model species within a single model by explicitly accommodating residual species correlations, which facilitates co‐occurrence hypothesis testing (Ovaskainen et al, 2010) and increases the precision of both individual species distributions and community metrics. However, most JSDMs typically do not accommodate imperfect detection (but see Hogg et al, 2021; Tobler et al, 2019). Failure to account for imperfect detection in detection–nondetection data can lead to biases in estimates of both species distributions and the effects of environmental drivers on species occurrence (MacKenzie et al, 2002).…”

“…Despite separate development of JSDMs that account for residual correlations and imperfect detection, only recently have approaches emerged that incorporate both of these complexities in JSDMs for large communities (Hogg et al, 2021; Tobler et al, 2019). Further, these approaches can become computationally intensive as both the number of spatial locations and species in the community increase, and no approaches exist that simultaneously incorporate species correlations, imperfect detection, and spatial autocorrelation, despite the well recognized impacts of ignoring these complexities.…”

Joint species distribution models with imperfect detection for high-dimensional spatial data

“…By jointly modeling species within a single model, JSDMs facilitate co-occurrence hypothesis testing (Ovaskainen et al, 2010) and increase precision of both individual species distributions and community metrics. However, JSDMs typically do not accommodate imperfect detection (but see Tobler et al 2019;Hogg et al 2021). Failure to account for imperfect detection when modeling detection-nondetection data can lead to biases in both species distributions and the effects of environmental drivers on species occurrence (MacKenzie et al, 2002).…”

“…Despite development of JSDMs, multi-species occupancy models, and their spatially-explicit extensions, only recently have approaches emerged that incorporate species correlations and imperfect detection in SDMs for large communities (Tobler et al, 2019;Hogg et al, 2021).…”