A method to estimate actual infrastructure‐induced mortality by integrating hierarchical sources of imperfect detection

Abstract Human infrastructures are among the most impactful wildlife threats. Although estimates of animal mortality by these structures exist over a given period, they typically do not account for several levels of imperfect detection (i.e. difference between recorded and true mortality). Consequently, true mortality rates may be severely underestimated, as well as their impact on populations and species. We present a hierarchical Bayesian latent‐state modelling framework that sequentially accounts for three processes that produce imperfect detection in estimating mortality abundance: the probability that a hit animal dies on the surveyed area (carcass location probability), the probability that the carcass remains on the surveyed area until the survey is conducted (carcass persistence probability) and the probability that the carcass is observed during the survey process (carcass observation probability). We employ a simulation study where we test the effects of variability in species characteristics, sampling design, latent‐state parameters and prior information on the ability of our model to estimate mortality abundance on roads as total number of roadkills. We then apply our framework to a case study to estimate the total number of roadkills per km in Mediterranean ecosystems while evaluating the cross‐efficiency of different sampling methods. Our framework accurately recovers the total number of roadkills from simulated census data for most simulation scenarios. We detected the highest disagreement between modelling outcomes and simulated data when variability in simulated carcass persistence probability was high and Bayesian priors were highly diffuse. In the case study, our results show notably high roadkill numbers (e.g. estimating 48.92 per km passerines based on 8.04 observed counts), along with substantial variation across different vertebrate groups. Furthermore, our case study confirms that walking and cycling surveys outperform driving surveys in carcass observation rate and provide complementary information, observing partially distinct sets of species and carcass sizes. Our modelling framework offers an efficient approach to estimate mortality rates for a wide range of taxa. Optimizing application requires extensive fieldwork for estimation and integration of imperfect detection. We provide a checklist to help managers assess when infrastructure‐related mortality can be assessed most robustly to prioritise conservation efforts.