The use, misuse and opportunities for structural equation modelling ( <scp>SEM</scp> ) in wildlife ecology

Abstract Technological advances in passive detectors like wildlife cameras and bioacoustics have immense potential to help community ecologists understand species interactions and food webs. However, while passive sensors are relatively easy to operate, the statistical tools needed to study complex causal pathways that constitute food webs have a steep learning curve. Here, we review analytical tools for assessing ecological interactions using observational datasets. We motivate the use of path analysis and structural equation modelling (SEM) by showing how they build on generalised linear mixed modelling and relate them to hierarchical models that account for detectability. Using a simulated dataset of wildlife observations from cameras, we compare the functionality and performance of the four dominant SEM packages available in R. The top performer was piecewiseSEM, with paths fit by the glmmTMB package, and zero‐inflation accounted for by mediator variables. The brms package was the second‐best performer. However, no SEM package possesses all the desirable functionalities, and they vary in terms of estimated bias, precision (standard errors) and computational time. We also compare the use of discrete versus continuous distributions for data derived from passive detectors, finding that the zero‐inflated negative binomial distribution (ZINB) is the most flexible and least biased for SEMs using count data. Synthesis and applications: The integration of SEM with observational wildlife datasets from passive detectors can lead to a step change in the scope, richness and robustness of insights about food webs, ecosystem functioning and conservation, but accounting for detectability remains a key hurdle.